Curator state

Recently surfaced

• Five ToolUniverse drug-discovery skills — five new (added 2026-08-16) — Drug Repurposing and Discovery directed pass. The drug repurposing MCP server seed query returned no uncatalogued, license-clear MCP server: the strongest hit, NovoMCP (Apache-2.0, 68 tools, NovoMCP/novomcp), is already catalogued as novomcp.md; the BioMCP / ChEMBL / openFDA wrappers it also surfaced are already covered (biomcp.md, chembl.md, openfda.md, openfda-mcp-server.md). The pass therefore drew again from the ToolUniverse skills queue (mims-harvard/ToolUniverse, Apache-2.0), whose skills/ listing was re-fetched in full this run — it has grown well past the ~68 previously recorded — cataloguing the five DRD-relevant directories with no existing coverage, each verified against its GitHub SKILL.md front-matter and body this run. tooluniverse-gene-liability (tooluniverse-gene-liability.md, DRD + MCB + Translational Medicine) — target safety scoring, the complement to the already-catalogued tooluniverse-drug-target-validation (efficacy): MyGene_query_genes resolution → five evidence dimensions (gnomad_get_gene_constraints, OpenTargets_get_biological_mouse_models_by_ensemblID, GTEx_get_median_gene_expression, OpenTargets_get_target_safety_profile_by_ensemblID, DepMap_get_gene_dependencies, plus OpenTargets_get_associated_drugs_by_target_ensemblID for precedent) → 0–25 points per dimension on T1–T4 evidence tiers → liability = 100 × points earned / available weight (dimensions with no data leave the denominator rather than scoring as safe) → a modulation-strategy recommendation (full / partial / transient / tissue-specific / deprioritise). Cutoffs carried onto the page: pLI ≥ 0.9 + LOEUF ≤ 0.35 = full 25 for constraint, embryonic lethality = full 25 for knockout phenotype. Note: the fetched front-matter showed only name + description (no disable-model-invocation), so that sentence was deliberately omitted from this page unlike the other four — re-check on the next DRD slot. tooluniverse-gpcr-structural-pharmacology (tooluniverse-gpcr-structural-pharmacology.md, DRD + Structural) — GPCRdb_list_proteins / GPCRdb_get_protein → GPCRdb_get_ligands (agonist / antagonist / inverse-agonist / biased-agonist classification) → GPCRdb_get_structures filtered by active/inactive/intermediate state with PDBePISA_get_interfaces + PDBePISA_get_assemblies → GPCRdb_get_mutations mapped onto the binding site (how orthosteric vs allosteric gets resolved) → SAbDab_search_structures / SAbDab_get_structure for antibody CDRs and interfaces. First GPCRdb, PDBePISA and SAbDab surface in the catalog — none of the three has a discrete entry. tooluniverse-mendelian-randomization (tooluniverse-mendelian-randomization.md, DRD + Translational Medicine) — EpiGraphDB_search_opengwas exact trait-label resolution → EpiGraphDB_get_mendelian_randomization (pre-computed MR-EvE estimates) or OpenGWAS_get_mr_instruments (custom two-sample) → effect direction/magnitude read against instrument quality via the MOE score → triangulation across IVW / MR-Egger / weighted median, bidirectional MR for reverse causation, and EpiGraphDB_get_genetic_correlations explicitly reported as not causation → optional drug-target follow-up. Two upstream scope exclusions carried onto the page: plain GWAS association lookups route to the GWAS skills, and it will not fit instruments from raw summary statistics the user supplies. tooluniverse-aging-senescence (tooluniverse-aging-senescence.md, DRD + MCB + Translational Medicine) — the catalog’s first geroscience entry: seven phases (query parsing → 12-hallmarks classification → OpenGenes_get_gene / OpenGenes_search_genes + gwas_get_snps_for_gene / gwas_search_associations → KEGG_get_pathway_genes / kegg_search_pathway + STRING → senolytics (dasatinib+quercetin, fisetin, navitoclax) and OpenTargets_get_associated_targets_by_disease_efoId → PubMed_search_articles → evidence-graded report separating correlative from causal/knockout evidence). Covers p16/CDKN2A, SASP, SA-β-gal, epigenetic clocks, telomere biology, longevity GWAS, centenarian genetics. OpenGenes is a new data source for the catalog. tooluniverse-enzyme-kinetics (tooluniverse-enzyme-kinetics.md, Chemistry + DRD + MCB) — EnzymeKinetics_calculate (operations michaelis_menten, inhibition) plus a bundled fit_michaelis_menten.py CSV script: initial-rate validation, substrate range must span Km, ≥5–7 points, enzyme concentration required for kcat; nonlinear MM regression returning Vmax/Km/R²/SSE with Lineweaver-Burk emitted for reference only; kcat/Km benchmarked against the ~10⁸–10⁹ M⁻¹s⁻¹ diffusion limit; competitive/uncompetitive/non-competitive assignment with Ki; R² ≥ 0.98 is the good-fit bar; flags substrate inhibition, out-of-range Km, missing [E], residual patterns. Upstream exclusion carried onto the page: it fits your data and is not a BRENDA lookup of published constants (cross-linked to brenda-database.md), and it is distinguished from the 4PL/Hill tooluniverse-dose-response.md. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read, all plugins already catalogued, no new entries.
• Medicare MCP + four bioSkills clinical-databases/ skills — five new (added 2026-08-16) — Translational Medicine directed pass, clearing two long-standing Deferred items. (1) Medicare MCP Server (openpharma-org/medicare-mcp) — catalogued medicare-mcp.md, tagged Translational Medicine. The 2026-07-26 blocker cleared: the GitHub license API now returns MIT with a LICENSE file in the repo root (the sibling fda-mcp cleared the same way on 2026-08-09, and this run confirms the org has licensed both). Correction to the original deferral note: there is no npm package — both medicare-mcp and the package.json name @openpharma-org/medicare-mcp-server 404 on the registry (checked 2026-08-16) — so the page installs from source (git clone → npm install → npm run build → register node /path/to/medicare-mcp/build/index.js), exactly as fda-mcp.md does, and it warns that the upstream README’s /path/to/medicare-mcp-server/… path does not match the repo’s own directory name medicare-mcp. package.json verified: v0.3.0, bin → build/index.js, engines.node >=16.0.0, deps @modelcontextprotocol/sdk ^1.11.4 + dotenv + express + node-fetch. Availability recorded as Beta (source-only install). Single medicare_info tool with 17 methods in four groups — provider/prescriber/hospital utilization (2013–2023), hospital quality and safety (star ratings, 30-day readmission and mortality, HAI, VBP, HCAHPS, compare), Part B ASP pricing + trends, and Part D formulary coverage + trends. Keyless (CMS public APIs). Upstream self-describes as “Unofficial”, quarterly CMS refresh cadence noted, and the page flags that provider-level Medicare data names individual clinicians. Cross-linked to cms-datagov-mcp.md (dataset catalog), cms-coverage.md (coverage determinations) and fda-mcp.md. (2–5) Four bioSkills clinical-databases/ skills (GPTomics/bioSkills, MIT) — the 12-skill category listing was re-confirmed and the four genuinely-distinct oncology/germline-interpretation skills catalogued, all tagged Translational Medicine: acmg-classification.md (bio-clinical-databases-acmg-classification, primary_tool requests) — ACMG/AMP 2015 as practised today: ClinGen SVI CSpec/VCEP rules, Tavtigian 2018/2020 Bayesian point system, Abou Tayoun 2018 PVS1 decision tree (AutoPVS1), Pejaver 2022 REVEL/BayesDel/VEST4 and Bergquist 2025 AlphaMissense calibrated PP3/BP4 bands, Walker 2023 SpliceAI DS_max ≥ 0.2 → PP3_Supporting plus SpliceVault, Brnich 2020 OddsPath for PS3/BS3, Whiffin 2017 max-credible-AF for BS1 and SVI BA1 > 5%, and AMP/ASCO/CAP 2017 Tier I–IV for somatic variants kept separate from germline P/LP/VUS/LB/B; tumor-mutational-burden.md (primary_tool cyvcf2) — Friends of Cancer Research harmonization with per-assay calibration (FDA 10/Mb ≈ 7.8 TSO500 ≈ 8.4 Oncomine TML), VAF ≥ 5% paired / ≥ 10% tumour-only, depth ≥ 100, gnomAD AF ≤ 0.5% germline filtering, normalization by scored Mb not nominal panel size, hypermutator tiering, blood TMB vs tissue, McGrail 2021 tumour-specific cutoffs, ESMO 2024 / FDA pembrolizumab 2020 reporting, and HLA-LOH (LOHHLA/DASH) + Luksza 2017 neoantigen-fitness integration; msi-detection.md (primary_tool MSIsensor-pro, tool_type: cli) — MSIsensor (paired) / MSIsensor-pro (tumour-only + cohort baseline) / MSIsensor-ct (cfDNA) / MANTIS (step-wise difference > 0.4) / mSINGS, with MSI-H ≥ 20–30% unstable loci (Bethesda ≥ 2/5 = 40%), ≥ 50 informative loci, purity ≥ 20%, ctDNA ≥ 3%, IHC/dMMR concordance as a required cross-check, and the POLE-exo-hypermutator confusion called out; somatic-signatures.md (primary_tool SigProfilerAssignment, also tagged Molecular and Cellular Biology) — COSMIC v3.4 (86 SBS / 11 DBS / 18 ID / 21 CN / 16 SV) via SigProfilerSuite, MutationalPatterns, MuSiCal mvNMF, SigNet and HRDetect, with the de-novo-extraction-vs-refit decision gated on cohort ≥ 50, stability gates (100 replicates, min ≥ 0.2 / avg ≥ 0.8), and etiology→action mapping (BRCA1/2 HRD → PARP, MMR-D/POLE → ICI, APOBEC3A, UV, tobacco, aflatoxin, 5-FU SBS17b, platinum, colibactin SBS88). All four carry a research-use-not-diagnostic framing and are mutually cross-linked (TMB ↔ MSI ↔ signatures as the three complementary reads of one somatic call set); the COSMIC-terms caveat is stated on the signatures page since the MIT skill licence does not extend to the signature catalogue. Deliberately not catalogued from this category: clinvar-lookup, dbsnp-queries, gnomad-frequencies, myvariant-queries, hla-typing, pharmacogenomics and variant-prioritization wrap data sources already catalogued — fold the bioSkills install path into those pages rather than duplicating. polygenic-risk is a genuinely-distinct fifth candidate deferred only by the soft cap (see Deferred). Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-fetched — 18 plugins listed, all already catalogued, zero diff; anthropics/claude-plugins-official/.claude-plugin/marketplace.json scanned — no new life-science entries (boltz already catalogued; fiftyone is a computer-vision dataset tool and was not pursued). Note for future sweeps: myvariant has no catalog page despite being named as an All-tagged example in the tagging guidance — worth checking whether MyVariant.info deserves one alongside mygene.md.
• NWB MCP Server + neuro-mcp — two new (added 2026-08-16) — Neuroscience directed pass. The headline result is that the standing NWB gap, reconfirmed on the 2026-06-05 and 2026-08-02 Neuroscience passes, is now closed. (1) NWB MCP Server (bjhardcastle/nwb-mcp-server, MIT via the GitHub license API, default branch main, PyPI 0.1.9 released 2026-06-07, 2 stars) — catalogued nwb-mcp-server.md, tagged Neuroscience. Exposes NWB files as a virtual SQL database via lazynwb, ten read-only tools (get_tables, get_table_schema, get_nwb_paths, preview_table_values, execute_query, get_active_source, use_local_source, use_dandiset_source, reset_active_source, nwb_file_search_code_snippet), with local, S3 and DANDI dandiset sources selectable at runtime. Install-path facts carried onto the page: uvx nwb-mcp-server needs no clone; the upstream README is written for VS Code Copilot Chat and publishes only an mcp.json command/args pair, so the Claude Code claude mcp add --transport stdio and Claude Desktop mcpServers forms are adapted from it and marked **Unverified —** rather than presented as documented; running the binary is labelled a one-shot boot check, not a service; the README’s underscore flag spelling (--root_dir, --glob_pattern) is followed verbatim; and every /path/to/… placeholder is annotated. Limits stated because they bite in practice: --infer_schema_length defaults to 1 file so heterogeneous collections get an incomplete schema, --max_result_rows defaults to 50 and table_element_limit to 500 elements so results truncate rather than stream, and lazynwb is pinned to the pre-release 1.0.0dev3. Positioned as the tabular side of NWB (units/trials/epochs/electrodes/subject metadata) and cross-linked to spikelab.md, spikeinterface-electrophysiology.md and neurosift.md for waveform analysis and DANDI semantic search. (2) neuro-mcp (AImplifier/neuro-mcp, BSD-3-Clause, PyPI 0.1.3 released 2026-07-24, MCP Registry io.github.AImplifier/neuro-mcp v0.1.3 active, 0 stars) — catalogued neuro-mcp.md, tagged Neuroscience + Translational Medicine. 54 tools in three layers: MNE-Python preprocessing (load_neuro/filter_neuro/resample_neuro/set_montage/set_reference/detect_bad_channels/run_ica/apply_ica/find_events/epoch_neuro/compute_psd/compute_erp/time_frequency/plot_*), source imaging from fetch_template_head through extract_label_timecourses, and a BIDS/EHR record store (subjects, versioned amend/void EHR entries, datasets, annotations, get_audit_log) plus NeuroII visualization exporting self-contained offline Plotly HTML. This is one of the few catalogued neuroscience servers that writes, so the page leads with that: DATABASE_URL (default SQLite at ~/.neuro-mcp/neuro_mcp.db) and BIDS_ROOT should point at scratch locations first, Postgres is the multi-user path via the [postgres] extra, and 3-D ESI rendering needs [viz3d]. Two honesty notes: the project markets itself as assisting clinicians but publishes no regulatory clearance and no medical disclaimer — the page supplies the research-grade-only framing upstream omits — and at 0 stars / v0.1.3 / one organization / no publication it is genuinely unproven, so the page tells readers to check any pipeline against hand-run MNE. Registration uses the environment’s interpreter (/path/to/envs/neuro-mcp/bin/python -m neuro_mcp), not a bare python, and both the Claude Code and Claude Desktop forms are shown. Framed against mne-eeg-tool.md (the skill teaches Claude to write MNE code; this server exposes MNE as callable tools). Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-fetched — 20 plugins, all catalogued, zero diff. Also evaluated and not catalogued: sensein/brainkb_mcp (no license, and its remote endpoint is documented as “future” — see Deferred), scitex-ai/scitex-dataset (already covered by scitex.md), braingeneers/SpikeLab and allenbrain-mcp (already catalogued), manncodes/neuronpedia-mcp (mechanistic interpretability of language models, not neuroscience — a recurring name-collision trap), and catalystneuro/mcp_read_images (a general vision server, not NWB). Standing Neuroscience gaps re-confirmed this pass: no MCP or Skill wraps the ABC Atlas abc_atlas_access S3 stack (allenbrain.md targets the older RMA API), no connectome-specific MCP exists, no calcium-imaging pipeline server (Suite2p/CaImAn) exists, and NWB conversion (neuroconv, NWB GUIDE, NWB Inspector) is still unwrapped — the new server is read-only query, not conversion.
• MCPmed org sweep + R bridge — three new (added 2026-08-15) — Molecular and Cellular Biology directed pass, executing the MCPmed full-org sweep that the 2026-08-15 Immunology pass banked for an MCB slot. Org listing pulled via the GitHub API (12 repos): GEOmcp ✅ and allenbrain-mcp ✅ already catalogued, STRINGmcp not yet linked from string-database-ppi.md (augment, see Deferred), PLSDBmcp still unlicensed (see Deferred), breadcrumbs/Cookiecutter-MCPmed/.github are templates and out of scope, paperscraperMCP and EMBL-EBI-Protein-mcp overlap existing literature/protein pages. Two genuinely new servers catalogued plus one utility. (1) UCSC Cell Browser MCP (MCPmed/UCSCCBmcp, BSD-3-Clause per repo LICENSE) — ucsc-cell-browser.md, tagged Molecular and Cellular Biology. Seven read-only tools over the public cells.ucsc.edu/dataset.json index (search_datasets, get_dataset_details, list_organisms, list_body_parts, list_diseases, list_projects, refresh_data), no API key, deps just httpx + mcp, Python 3.8+. Two upstream README errors corrected on the page rather than copied: pip install ucsc-cell-browser-mcp does not work — the PyPI package 404s (checked 2026-08-15), and the README’s cd ucsccbMCP is a typo for UCSCCBmcp; the page therefore ships the source install (pip install -e .) plus which ucsc-mcp → literal claude mcp add and claude_desktop_config.json snippets, and states explicitly that running ucsc-mcp is a boot check only. License discrepancy recorded, not papered over: repo LICENSE says BSD-3-Clause, pyproject.toml says MIT. Positioned as discovery-only (no matrix download) and cross-linked to cellxgene-census.md, geo-database.md, scanpy.md. (2) PBMCpedia MCP (MCPmed/PBMCpediaMCP, BSD-3-Clause, default branch master, TypeScript) — pbmcpedia.md, tagged Immunology and Microbiology + Molecular and Cellular Biology. Over the Saarland CCB harmonized PBMC scRNA-seq atlas (unified reference mapping + cell-type annotation, with TCR/BCR and CITE-seq ADT layers; NAR doi:10.1093/nar/gkaf1245). The README documents no tools at all — the six were read directly from server.ts this run and the page says so: getMetaData, getExpressionPerGene, getDEGs, getDEperCellType, getPathways, getAntibodyChains, with age-group/sex/disease/cell-type (fine and broad) filters and pagination. No npm package — clone + npm install. The cwd trap is called out: claude mcp add has no --cwd flag, so the stdio entry in ~/.claude.json must be hand-edited to add "cwd" or npm start runs in the wrong directory and the server never launches; HTTP alternative documented at the server.ts default port 3002 with the note that Desktop has no native HTTP transport. (3) mcptools (posit-dev/mcptools, MIT, CRAN 1.0.1 published 2026-07-27, Posit Software PBC / Simon Couch) — mcptools.md, tagged General-Purpose Utilities. This closes a structural gap: a Bioconductor MCP server search confirmed no Bioconductor-specific MCP server exists, yet the catalog already carries a cluster of R/Bioconductor-based skills (deseq2-differential-expression.md, cytometry-qc.md, gating-analysis.md, clustering-phenotyping.md, compensation-transformation.md, cytometry-differential-analysis.md, hashing-demultiplexing.md) with no catalogued way to reach an R runtime. install.packages("mcptools") + claude mcp add -s "user" r-mcptools -- Rscript -e "mcptools::mcp_server()"; .Rprofile mcp_session() snippet for live-session attachment, with the working-directory matching rule. Two honesty notes on the page: the default surface is only list_r_sessions()/select_r_session() — real capability requires registering ellmer::tool() wrappers — and the Desktop JSON is marked **Unverified —** because upstream documents only the claude mcp add form. Security caveat carried: a run_r_code() tool is arbitrary code execution with the session’s filesystem and credential access (“for the brave” upstream). Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-fetched — 21 plugins, all catalogued, zero diff; MCP Registry search=single-cell returned zero servers. Standing MCB gap confirmed this pass: no CRISPR / gRNA-design MCP server or Skill exists anywhere — see Deferred.
• Four bioSkills alignment/ skills — four new (added 2026-08-15) — Integrative Structural and Computational Biology directed pass. The gap this closes: the alignment/ category of GPTomics/bioSkills (MIT) had never been fetched, and the catalog had no dedicated MSA-construction, alignment-trimming, alignment-statistics or structural-alignment surface at all — MAFFT appeared only inside phylogenetics.md (as one step of a tree-building skill), TM-align/Foldseek only inside protein-mcp-server.md and foldseek-structural-search.md, and trimAl/ClipKIT/BMGE nowhere. The 65-directory root listing was re-fetched this run and the 7-skill alignment/ listing captured in full (alignment-io, alignment-trimming ✅, msa-parsing, msa-statistics ✅, multiple-alignment ✅, pairwise-alignment, structural-alignment ✅); the four with no existing coverage were catalogued as a mutually cross-linked build → trim → measure → structural chain, all tagged Integrative Structural and Computational Biology (the three sequence-side ones also Molecular and Cellular Biology). structural-alignment (structural-alignment.md, bio-alignment-structural, primary_tool Foldseek — this closes the standing “no standalone TM-align/FoldMason” gap recorded under Deferred since 2026-08-01: Foldseek 8+ 3Di search at ~10³–10⁶ structures/s, TM-align 20220412+ single-chain, US-align 20231222+ for multi-chain/RNA/DNA with -mm 1 -ter 0, Foldseek-Multimer, DALI, FATCAT, CE, Bio.PDB.Superimposer, and FoldMason 1+ for structural MSA plus T-Coffee Expresso / 3D-Coffee hybrids; score bands carried onto the page — TM-score >0.5 same fold / >0.8 equivalent topology / <0.2 random, with a length-aware Gumbel p-value below 60 residues (Xu & Zhang 2010); DALI Z >20 definite / 8–19 probable / 2–8 candidate / <2 not significant; RMSD <2 Å over >100 residues strong, <1.5 Å excellent; lDDT >0.6 (Mariani 2013). The load-bearing sequence-identity ladder — ≥40% plain DP, 25–40% MMseqs2/jackhmmer/HHsearch, 15–25% profile–profile or Foldseek, <15% structural/pLM aligners only, with a stated exception down to ~12–15% under strong structural constraint. Headline install-blocking warning: mask residues with pLDDT < 70 before Foldseek indexing — their coordinates encode as random 3Di letters and contaminate hits below TM ≈ 0.4, i.e. feeding a full-length AlphaFold model with disordered tails into a structural search manufactures confident nonsense; plus the RMSD-is-not-a-property and Superimposer-does-not-solve-correspondence boundaries). multiple-alignment (multiple-alignment.md, bio-alignment-multiple, primary_tool MAFFT; MAFFT 7.520+ / MUSCLE 5.1+ / ClustalOmega 1.2.4+ / T-Coffee 13+ / PAL2NAL 14+; the MAFFT --auto boundaries made explicit — <200 L-INS-i, 200–500 FFT-NS-i --maxiterate 2, 500–2,000 FFT-NS-2, 2,000–50,000 single-pass, >50,000 PartTree — and per-tool capacity L-INS-i ~200 / MUSCLE5 PPP ~1,000 / super5 100,000+ / ClustalOmega 190,000 benchmark; the twilight-zone table >40% any tool / 25–40% iterative + GUIDANCE2 / 20–25% profile–profile / <15–20% switch to structural alignment; six-stage workflow gated on a BLAST E < 1e-5 homology check; codon routing table (clean orthologs → MAFFT-protein + PAL2NAL, indel-rich paralogs → PRANK +F codon, frameshifts/pseudogenes → MACSE v2, mixed → OMM_MACSE, HyPhy-grade → pre-msa.bf/post-msa.bf) behind the sharp warning that nucleotide MSA tools break reading frames and yield systematically incorrect dN/dS and false-positive selection even on clean simulated data (Fletcher & Yang 2010). API fact carried onto the page: Bio.Align.Applications was removed in BioPython 1.86, so the aligners must be driven via subprocess). alignment-trimming (alignment-trimming.md, bio-alignment-trimming, primary_tool ClipKIT; ClipKIT 2.1+ / trimAl 1.4+ / BMGE 1.12+ / Divvier 1.01+ / HMMcleaner; BMGE -h 0.5 default with 0.4 deep-phylogenomics and 0.6 shallow, -g 0.2, -b 5, TCS retain ≥5 on the 0–9 scale; the headline >60% column-retention rule — removing more than 40% of columns means the mode is too aggressive — plus the mandated before/after tree-topology sensitivity analysis and the column-index mapping requirement for any site-specific downstream claim; the column-trimmers-vs-HMMcleaner distinction, since HMMcleaner removes segments within individual sequences rather than columns across all of them). msa-statistics (msa-statistics.md, bio-alignment-msa-statistics, primary_tool Bio.Align, pure Python — BioPython 1.83+ / NumPy 1.26+; the four percent-identity denominators PID1/PID2/PID3/PID4 with PID4 recommended for orthologs, and the reproducibility headline that they vary by up to 11.5% on the same alignment and 22% once algorithms differ, with the worked 500-residue-vs-100-residue-fragment case giving PID2 ≈ 84% against PID4 ≈ 27%; Capra–Singh JSD conservation (AUC ≈ 0.94, Top-30 ≈ 0.75) vs entropy/majority vs ConSurf rate4site; IC via log₂(4) − H for DNA and KL against Robinson & Robinson background for protein; quality gates >50% gaps in >30% of columns → re-align, <25% mean protein PID → GUIDANCE2 or structural alignment; Karlin–Altschul lambda calibration differences (NCBI BLAST 0.3176, FASTA/SSEARCH per-query, HMMER Forward-calibrated, Bio.Align not implemented) giving ~2% bit-score spread; and the silent-failure trap that substitution_matrices.load('BLOSUM62') returns a NumPy-backed Array, not a dict, so .get((c1,c2), 0) always returns 0 without raising). Install-path facts verified this run: every prerequisite has a clean bioconda one-liner and each satisfies the skill’s stated minimum — foldseek 8.ef4e960, tmalign 20240303, usalign 20241201, foldmason 2.7bd21ed, mafft 7.525, muscle 5.3, clustalo 1.2.4, t-coffee 13.46.0, clipkit 2.14.0, trimal 1.5.1, bmge 1.12, divvier 1.01 (all checked 2026-08-15 via the anaconda.org API). Three gaps marked **Unverified —** rather than hand-waved: HMMcleaner is CPAN-only (cpanm Bio::MUST::Apps::HmmCleaner) with no pinned version upstream, the codon extras (pal2nal/prank/macse) were not version-checked, and ConSurf/rate4site has no upstream install path. DALI is documented as an external web service, not a conda package — with the note that uploading an unpublished structure sends it off the machine. Cross-linked into foldseek-structural-search.md, protein-mcp-server.md, geometric-analysis.md, structure-validation.md, phylogenetics.md, covariation-analysis.md, ncrna-search.md, blast.md, biopython.md. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-fetched — 20 plugins returned, all already catalogued; MCP Registry searches for structure (returns unrelated app-building/CRM servers — a name-collision trap worth noting) and protein (only cyanheads/protein-mcp-server and pipeworx-io/protein-atlas, both catalogued) yielded zero new servers; a MCP server protein structure prediction WebSearch returned only already-catalogued or already-deferred candidates (Augmented-Nature AlphaFold/PDB/STRING servers, chimerax-alphafold-mcp, boltz_mcp, zeinab-sheikhi/mcp-alphafold). Standing structural gaps re-confirmed: still no cryo-EM (RELION/cryoSPARC) MCP or Skill, and still no RNA 3D structure-prediction wrapper.
• CAN-IMMUNE + Blatant-Why — two new (added 2026-08-15) — Immunology and Microbiology directed pass. (1) CAN-IMMUNE (sanjaysgk/can-immune-mcp, MIT, MCP Registry io.github.sanjaysgk/can-immune v1.0.0 published 2026-08-10, status active) — catalogued can-immune.md, tagged Immunology and Microbiology + Drug Repurposing and Discovery + Translational Medicine. Hosted Streamable-HTTP MCP over the Li/Purcell Lab (Monash) cancer neoantigen catalogue: ~4,485,395 unique mutant peptides from COSMIC + DepMap/CCLE, 12 read-only tools (database_overview, search/fetch, search_genes, get_gene, get_gene_mutations, search_cell_lines, get_cell_line, top_genes_in_cell_line, list_tissues, list_cancer_types, get_mhc_binding), MHC-I binding predicted with NetMHCpan-EL via the IEDB API on 25-mer windows centred on the mutated residue, plus a shared-neoantigen view across tissues/cell lines. One-line install, no key: claude mcp add --transport http canelib https://canelib.erc.monash.edu/mcp; Claude.ai via custom connector; Desktop needs an mcp-remote proxy (marked Unverified — not documented upstream). Two hostnames in circulation and both recorded on the page: endpoint/docs at canelib.erc.monash.edu, browsable UI at canelib.erc.monash.edu.au. Anonymous access at a light rate; Google OAuth 2.1 raises quota with academic domains (.edu, .edu.au, .ac.*) highest. Caveats noted: repo is new (2 commits, 0 stars) though the service is live, no database publication locatable despite a “please cite” request, COSMIC/DepMap terms govern the underlying data even though the server code is MIT, and binding scores are predictions not measured immunogenicity. (2) Blatant-Why (BY) (001TMF/blatant-why, MIT, 104 stars, default branch master, npm blatant-why v0.1.0-beta published 2026-04-03, bin is by) — catalogued blatant-why.md, tagged Immunology and Microbiology + Drug Repurposing and Discovery + Integrative Structural and Computational Biology. Antibody/nanobody/de-novo-binder design campaign scaffold for Claude Code: 19 skills, 13 /by:… slash commands, 21 agents, 11 bundled MCP servers (pdb, uniprot, sabdab, screening, tamarind, cloud, adaptyv, campaign, research, local_compute, knowledge), driving BoltzGen (Ab/Nb), PXDesign (de novo binders) and Protenix (structure/confidence), with ipSAE scoring, liability scan, developability and diversity selection, and a JSON knowledge store for cross-session campaign memory. Install-path facts carried onto the page: it is a project scaffold, not a marketplace plugin — mkdir my-campaign && cd my-campaign && npx blatant-why init && claude is the only path, /plugin marketplace add will not find it, skills and commands are project-scoped so claude must be run from the campaign directory, prerequisites are Node 18+/Python 3.11+/uv/Claude Code, and Claude Desktop and Claude.ai are unsupported. Optional third-party spend documented (Tamarind free tier 10 jobs/month, RunPod ~$0.40–$2.50/hr, Adaptyv lab submission triple-gated behind /by:approve-lab); local GPU mode needs an NVIDIA CUDA GPU plus PROTEUS_FOLD_DIR/PROTEUS_PROT_DIR/PROTEUS_AB_DIR and no keys. Model weights themselves stay out of scope; cross-linked to adaptyv.md, pdb.md, uniprot.md, boltz.md. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-fetched at the corrected .claude-plugin/ path (the bare-root URL now 404s — use .claude-plugin/marketplace.json) — 21 plugins, all catalogued including cortellis and adisinsight; anthropics/claude-plugins-official/.claude-plugin/marketplace.json scanned, no new life-science entries (boltz already catalogued); Claude Science connectors-and-skills re-fetched, zero diff — all 17 featured connector source groups and all research skills (literature review, indication dossier, AlphaFold2, Boltz-2, Chai-1, ESMFold2, OpenFold3, ProteinMPNN/LigandMPNN/SolubleMPNN, DiffDock, ESM-2, Evo 2, Borzoi, scGPT, scvi-tools) are already catalogued. Standing gaps re-confirmed this pass: no MCP server or Skill wraps IEDB itself (its Query API and MHC-I/II + B-cell prediction tools remain reachable only indirectly, e.g. through can-immune.md and tooluniverse-*), and no MCP wraps BV-BRC (14M+ pathogen genomes, 33 analysis services, open API and p3-* CLI — the cleanest unfilled microbiology target). MCP Registry searches for epitope, antibody and microbiome all returned zero servers.

• ALKYL — one new (added 2026-08-15) — Chemistry directed pass (cheminformatics MCP server 2026 retrosynthesis Claude seed query → GitHub). ALKYL (Kdevos12/ALKYL, MIT via the GitHub license API, default branch master, plugin.json v1.0.0, 6 stars, 42 commits, last pushed 2026-03-15) — catalogued alkyl.md, tagged Chemistry + Drug Repurposing and Discovery + Integrative Structural and Computational Biology. The catalog’s first single-plugin computational-chemistry workbench: 27 skill directories read from the repo (cheminformatics rdkit/openbabel/daylight-theory/chem-brainstorm/synkit; MD+structure ase/mdanalysis/force-fields/coarse-grained/homology-modeling; QM qm-dft/organic-mechanisms; drug discovery docking/fbdd/free-energy/binding-kinetics/pharmacophore; design+ML generative-design/mmpa/uncertainty-qsar/active-learning/deepchem/torchdrug/pepflex; utilities py3Dmol/lit-rescue/nextflow), 22 standalone RDKit scripts, and four keyless bundled MCP servers (bioRxiv, ChEMBL, ClinicalTrials.gov, PubMed). Five install-path corrections carried onto the page, all of which would otherwise break a naive reader: (1) the upstream README’s clone URL is the placeholder https://github.com/YOUR_USERNAME/alkyl — the page substitutes https://github.com/Kdevos12/ALKYL and warns the created directory is capitalised ALKYL; (2) the repo ships .claude-plugin/plugin.json but no marketplace.json, so /plugin marketplace add Kdevos12/ALKYL will not resolve and the page says so explicitly — clone + bash alkyl.sh install is the only documented path; (3) the clone is the installation — alkyl.sh install only appends a block to the global ~/.claude/CLAUDE.md between <!-- ALKYL-START -->/<!-- ALKYL-END --> markers and skills are never copied to ~/.claude/skills/, so moving or deleting the clone silently breaks every skill (the single most likely failure mode; the page tells readers to clone somewhere permanent); (4) bash alkyl.sh venv (RDKit + pytest into a .venv/ in the clone) is optional and needed only for the standalone scripts, while bash alkyl.sh status is a one-shot check, not a service to leave running; (5) Claude Desktop is not supported — both mechanisms (~/.claude/CLAUDE.md, local script execution) are Claude Code-only. Also recorded: plugin.json advertises “23 domain skills” against 27 actual directories (the README merges torchdrug+pepflex), the heavy engines (ORCA, xTB, GROMACS, OpenMM, MODELLER, AutoDock Vina) are optional and not installed so a skill needing one fails until it is on PATH, the global-CLAUDE.md write means the chemistry context loads in every Claude Code session on the machine, and the ADMET/pKa scripts are estimators rather than validated models. Cross-linked to chemgraph.md (actually executes simulations) and xtb-mcp-server.md (input-deck prep only). Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-fetched — 15+ plugins returned (10x-genomics, pubmed, biorender, synapse, wiley-scholar-gateway, biorxiv, clinical-trials, chembl, owkin, open-targets, single-cell-rna-qc, instrument-data-to-allotrope, nextflow-development, scvi-tools, clinical-trial-protocol), all already catalogued; no new entries. Also evaluated and not catalogued this run: NovoMCP/novomcp and the CDXML/ChemDraw toolkit (both already have pages), s20ss/mcp_rdkit (already deferred as redundant against three existing RDKit entries), plus SYNTHIA Agentic AI, ghutchis/chem-skill and cyanheads/pubchem-mcp-server (all three newly recorded under Deferred with reasons). The spectra-prediction gap remains open — this pass re-confirmed no OSS MCP/Skill wraps SIRIUS, CFM-ID or NMR prediction.

• Four ToolUniverse regulatory/clinical-trial skills + the FDA MCP Server — five new (added 2026-08-09) — Drug Repurposing and Discovery directed pass. The mims-harvard/ToolUniverse skills/ listing was re-fetched this run and three previously-unnoticed DRD-relevant directories surfaced (tooluniverse-drug-mechanism-research, tooluniverse-drug-regulatory, tooluniverse-clinical-guidelines) alongside the two clinical-trial skills already on the deferred shortlist; four were catalogued (Apache-2.0, all four carry disable-model-invocation: true upstream so each page tells the reader to invoke explicitly). tooluniverse-drug-mechanism-research (tooluniverse-drug-mechanism-research.md, DRD + MCB + TM — the catalog’s first dedicated mechanism-of-action tracing surface; 8 steps drug resolution → primary target → off-target → pathway → FDA label → pharmacogenomics → literature → integration over OpenTargets_get_drug_mechanisms_of_action_by_chemblId / ChEMBL_get_drug_mechanisms / ChEMBL_get_target_activities / STRING_get_interaction_partners / KEGG_get_gene_pathways / DailyMed / PharmGKB / PubMed; the two rules carried onto the page — nanomolar affinity = primary target, micromolar = off-target at clinical dose, and multi-target convergence on one pathway means that pathway is the mechanism while targets in different pathways must be reported separately; four-tier evidence hierarchy T1 FDA labels + CPIC Level A + FDA PGx biomarkers / T2 ChEMBL mechanisms with references and affinity data / T3 database MOA and pathway entries / T4 PubMed–EuropePMC articles, with the page noting a T3-only mechanism is a database assertion not a measurement, and that ChEMBL mixes assay types so the nM/µM split is triage not a measured selectivity ratio). tooluniverse-drug-regulatory (tooluniverse-drug-regulatory.md, DRD + TM — no prior UNII/Orange-Book/exclusivity skill surface; 8 phases over FDAGSRS_* substance identity, RxClass_* ATC/EPC classification, FDA_OrangeBook_search_drug / _check_generic_availability / _get_te_code / _get_approval_history / _get_patent_info / _get_exclusivity, DailyMed_parse_* label parsing, trials, FAERS, literature; the constants carried onto the page — pathway definitions 505(b)(1) full package vs 505(b)(2) partial reliance vs ANDA bioequivalence-only, TE codes AB = substitutable vs BX/absent = not established, exclusivity durations NCE 5 y / ODE 7 y / PED +6 mo / NP-M variable, and the load-bearing heuristic that generic entry needs both patent expiry AND an approved ANDA, and an exclusivity code can block generics after every patent has expired — i.e. reading a patent expiry as a generic-entry date is the named error; jurisdiction must be named in every report, and clinical-trial AE rates kept separate from post-market signals. Upstream scopes out DDI, pharmacogenomics, MOA and repurposing to sibling skills, and the page says so; it also warns the EMA side is thinner than the FDA side and that Orange Book listings are not a freedom-to-operate opinion). tooluniverse-clinical-trial-matching (tooluniverse-clinical-trial-matching.md, DRD + TM — the distinguishing feature is three registries not one: ClinicalTrials.gov + EU CTIS + ISRCTN, so a CT.gov-only search silently misses EU/UK trials; 10 steps with 10 trials per batch characterization; evidence tiers T1 FDA-approved/guideline → T2 Phase III → T3 Phase I/II → T4 computational/preclinical; Trial Match Score 0–100 decomposed on the page as molecular match 0–40 (exact variant 40 / gene-level 30 / pathway 20 / none 10 / explicitly excluded 0) + clinical eligibility 0–25 (all 25 / most 18 / some 10 / ineligible 0) + evidence strength 0–20 (FDA-approved 20 / Ph III 15 / Ph II 10 / Ph I 5) + trial phase 0–10; safety framing is explicit — this is a shortlist for a clinician, not a clinical recommendation — with the sharp caveat that eligibility is scored by parsing free-text criteria, which is where registry entries are least structured, so a “criteria met” call is a hypothesis, while an excluded call is the component worth trusting most because exclusion criteria are stated explicitly). tooluniverse-clinical-trial-design (tooluniverse-clinical-trial-design.md, DRD + TM — six design dimensions (endpoint / population / comparator / effect size / duration / regulatory pathway) via six parallel research paths; feasibility score ≥75 HIGH, 50–74 MODERATE, <50 LOW with weights patient availability 30% / endpoint precedent 25% / regulatory clarity 20% / comparator feasibility 15% / safety monitoring 10%; endpoint grading A (regulatory acceptance, multiple precedents) → D (novel, unvalidated); 14-section markdown report. The page’s load-bearing correction: upstream states the skill is precedent-based reasoning, not first-principles math, so its “20% ORR improvement ≈ 100 patients/arm, 50% ≈ 30/arm” figures are order-of-magnitude anchors and the page tells readers not to put them in a protocol — routing the real calculation to power-and-sample-size.html and group-sequential/adaptive features to adaptive-designs.html; it also flags that Open Targets prevalence and gnomAD/COSMIC/ClinVar biomarker frequencies carry ascertainment bias straight into the eligible-population number, usually the most optimistic figure in the report). FDA MCP Server (OpenPharma) (fda-mcp.md, DRD + TM) resolves the candidate deferred 2026-07-26: the GitHub license API now returns MIT with a LICENSE file, clearing the sole stated blocker. Unique surface confirmed — Orange Book patent/exclusivity + Purple Book biosimilar interchangeability (47,486 drug products / 21,126 patents / 2,444 exclusivity periods / 2,168 biologics per upstream), neither covered by the catalogued openfda.md or openfda-mcp-server.md; single fda_info tool dispatched by method (lookup_drug, search_orange_book, get_therapeutic_equivalents, get_patent_exclusivity, analyze_patent_cliff, search_purple_book, get_biosimilar_interchangeability), keyless. Three install-path corrections carried onto the page: (1) the original deferral note’s “npm fda-mcp” is wrong — no npm package exists, both fda-mcp and the package.json name @openpharma-org/fda-mcp-server 404 on the registry, so the page builds from source (git clone → npm install → npm run build, Node ≥ 18, build: "tsc", entry build/index.js per main/bin); (2) the upstream README’s registration path /path/to/fda-mcp-server/build/index.js does not match the repo’s own directory name fda-mcp, which the page corrects and annotates with the pwd substitution; (3) both the Claude Code claude mcp add --transport stdio form and the Claude Desktop mcpServers JSON block are shown, and the node build/index.js step is labelled a one-shot Ctrl-C verification rather than a service to leave running. Availability recorded Beta (source-only install, repo pushed 2026-03-10, 3 stars); the page also warns that patent-cliff output is a research aid rather than a legal or commercial opinion, that cached datasets age between refreshes, and that the single-tool-with-methods design means naming the method explicitly in the prompt when a query comes back empty. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read — 20+ plugins returned, all already catalogued; no new entries.
• Five bioSkills clinical-biostatistics/ skills — five new (added 2026-08-09) — Translational Medicine directed pass, clearing the top-priority deferred queue item (GPTomics/bioSkills, MIT). The 12-directory clinical-biostatistics/ listing was re-fetched this run and confirmed unchanged since the 2026-08-02 audit (adaptive-designs ✅, bayesian-trials, categorical-tests, cdisc-data-handling, effect-measures, logistic-regression, missing-data-sensitivity ✅, multiplicity-graphical, power-and-sample-size, subgroup-analysis, survival-analysis, trial-reporting ✅). Catalogued the five strongest with no existing catalog coverage, as a mutually cross-linked confirmatory-trial statistics set: cdisc-data-handling (cdisc-data-handling.md — the catalog’s first CDISC/SDTM/ADaM surface anywhere; pyreadstat 1.2+ / pandas 2.1+ / numpy 1.26+ with R admiral/metacore/metatools/xportr for the derivation route; SDTM domain joins onto the DM spine, ADSL/BDS/OCCDS/ADTTE architecture with traceability, SUPPQUAL/NSV handling, Define-XML 2.1 generation, Pinnacle 21 4.0+ and open-source CORE validation; the install-blocking constants carried onto the page — XPT v5’s 8-character variable-name and 200-character value limits (the silent-truncation trap xportr exists to prevent), the ADY = (ADT − TRTSDT) + 1 no-day-zero convention, TEAE post-treatment windows that are sponsor-defined and modality-dependent (28–30 days small molecules per ICH E2A, 60–90 days biologics/mAbs, indefinite for cell and gene therapy), Define-XML 2.1 required for studies starting ≥ 15 March 2023, and Dataset-JSON v1.1 (Dec 2024) as the emerging XPT replacement per an FDA April 2025 notice — the page tells readers to confirm against the current Study Data Technical Conformance Guide rather than trust the page. Pricing row separates the MIT skill from CDISC membership requirements and commercial Pinnacle 21 Enterprise), power-and-sample-size (power-and-sample-size.md, statsmodels ≥0.14 / scipy ≥1.12 / numpy ≥1.26 / pandas ≥2.1 plus R pwr/gsDesign/gsDesign2/rpact/presize/npsurvSS/nph/simtrial/PowerTOST/clusterPower/ratesci/nlme; the headline δ ≥ 1.5 × MCID rule (Norman 2003 — δ = MCID leaves no margin for sampling variation) and the MCID-is-not-δ conflation as the named primary error; FDA 2016 NI M1/M2 double discount M2 ≤ 0.5 × historical M1 and M2 ≤ 0.5 × MCID at one-sided α = 0.025; Schoenfeld 1981 under-estimates by 20–50% under non-proportional hazards (Lin 2020 NPH Working Group) → Lakatos 1988 or simulation for delayed-effect/IO designs; Schuirmann TOST with the 90% CI inside (0.80, 1.25) FDA 1992 bioequivalence standard; dropout inflation n/(1−q), Senn 2013 stratification gain n/(1−r²), cluster design effect 1 + (m−1)·ICC; continuity correction wastes ~10% of sample size per D’Agostino 1988 Am Stat 42:198; CHW weighting required for unblinded SSR and the Mehta–Pocock promising zone at CP ≈ 30–80%; written to satisfy CONSORT 2025 item 16a), bayesian-trials (bayesian-trials.md, RBesT 1.7+ primary / OncoBayes2 0.8+ / BOIN 2.7+ / dfcrm 0.2-2+ / escalation 0.1+ / trialr 0.1.6+ / bayesDP / psborrow2 / bhmbasket / c212; BOIN target DLT 30% with escalate below 0.6× and de-escalate above 1.4× target, 12-patient lowest-dose futility floor, CRM indifference halfwidth 0.05 (Lee–Cheung 2009), EWOC P(dose > MTD) ≤ 0.25 (Babb–Rogatko–Zacks 1998); MAP prior ESS 20–80% of the new control (Schmidli 2014) with robust-mixture weight 0.1–0.3; EXNEX 0.5/0.5 Neuenschwander 2016 baseline with a required sensitivity sweep over 0.1/0.3/0.5/0.7/0.9; I-SPY 2 graduation at posterior predictive probability ≥ 0.85; Stan diagnostics R-hat < 1.01 (Vehtari 2021) and ESS > 1000 per chain; the page’s load-bearing framing is that borrowing weights are the design, so a single-weight analysis is not interpretable. Paediatric power-prior γ = 0.3–0.6 is flagged as a working convention the FDA January 2026 draft explicitly does not prescribe — and the draft is labelled a draft. Also notes the Stan C++ toolchain / multi-minute first install), multiplicity-graphical (multiplicity-graphical.md, gMCP 0.8.16+ / graphicalMCP 0.2+ / gatekeeping / multcomp / multxpert with statsmodels 0.14+ multipletests as the Python fallback; the four-step decision tree rendered on the page — confirmatory→FWER vs exploratory→FDR, strict priority→fixed-sequence/chain graph, then ≤3 endpoints unknown dependence→Holm, ≤3 with proven positive correlation→Hochberg/Hommel after a PRDS check, ≥4 with a primary/key-secondary/subgroup hierarchy→graphical gMCP, multiple families→parallel or serial gatekeeping; the sharp fourth branch, endpoints that can move in opposite directions (LDL vs HDL, efficacy vs safety) → Holm only, because PRDS fails and Hochberg is anti-conservative there; Bonferroni 30–50% power loss at ~10 tests under positive dependence (Sarkar 1998), Hommel 1–3% over Hochberg (Hommel 1988 Biometrika), subgroup α budget ≤ 20% of total (Dane 2019 EFSPI); anchored to the FDA Multiple Endpoints final guidance, October 2022, with the graph required in the SAP pre-unblinding), subgroup-analysis (subgroup-analysis.md, statsmodels 0.14+ primary / scipy 1.12+ / numpy 1.26+ / pandas 2.1+ / matplotlib 3.8+ / scikit-learn 1.4+ / econml with R grf/policytree/causalToolbox/personalized/SIDES/stepp/gMCP/partykit/RBesT/brms; scenario→method table (pre-specified confirmatory→interaction term + gMCP per the EMA 2019 “assessment subgroup”; stratified randomisation→CMH or strata-as-covariates because Kahan–Morris 2012 show ignoring strata biases SEs upward and loses power; many subgroups→Bayesian shrinkage; exploratory→STEPP/SIDES/causal forests labelled hypothesis-generating); the ~4× sample size interaction penalty (Brookes 2004) that makes most subgroup analyses underpowered by construction, 0.5σ noteworthy-heterogeneity benchmark (Dane 2019 EFSPI), ≤ 20% of total α for exploratory subgroups (EMA 2019), n ≥ 5 per cell for CMH and Breslow–Day unreliable below 5 strata, shrinkage prior tau ~ HalfNormal(0, 0.1) conservative / (0, 0.5) regulatory-tolerant swept over 0.1–0.5, causal-forest honest-split fraction 0.5 with tune.parameters = "all", RERI sign reading; credibility assessed against the Sun BMJ 2012 11-criterion checklist and EMA 2019; the headline rule that a subgroup-specific p-value is not evidence of a differential effect — the interaction test is). A sixth candidate, effect-measures, was also catalogued (effect-measures.md, statsmodels 0.14+ / marginaleffects 0.0.13+ Python and 0.20+ R / ratesci / exact2x2 / riskCommunicator / RobinCar / MASS / VGAM; OR/RR/RD/HR/NNT with Wilson, Newcombe, MOVER, Miettinen–Nurminen, exact-unconditional, profile-likelihood and Bender NNT intervals replacing Wald-on-a-ratio-scale; marginal vs conditional estimands per the FDA 2023 covariate-adjustment guidance, g-computation via marginaleffects with HC3 SEs; Hauck–Donner detection via VGAM::hdeff(); the two rules carried onto the page — OR and HR are non-collapsible so an adjusted conditional OR is not the same quantity as the unadjusted one even absent confounding, and an NNT interval runs through infinity when the risk-difference interval spans zero and must be reported as two intervals) — six pages this run rather than five, because effect-measures is the reporting layer the other five feed and splitting it across runs would have shipped an incomplete set. Deliberately not catalogued: survival-analysis (overlaps scikit-survival.md — fold the bioSkills path in there instead), logistic-regression and categorical-tests (generic statistics; belong on the General-Purpose Utilities shelf if anywhere). All six tagged Translational Medicine only. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read — 20+ plugins returned, all already catalogued (positions 18–20 medidata, consensus, cortellis all covered); no new entries.

• Five awesome_cognitive_and_neuroscience_skills skills — five new (added 2026-08-09) — Neuroscience directed pass, drawing the second batch from the collection audited 2026-08-02 (HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills, MIT, default branch master). The 50-directory skills/ listing was re-fetched this run and confirmed unchanged since the audit. Catalogued the five strongest with no existing catalog coverage, spanning four previously-empty Neuroscience surfaces: neural-population-analysis-guide (neural-population-analysis-guide.md — the catalog’s first population-geometry/latent-dynamics surface; PCA vs GPFA vs dPCA vs jPCA decision tree with t-SNE/UMAP explicitly ruled out for quantitative claims, per-method recording minima (PCA 30+ neurons; GPFA 50–100 neurons / 20–50 ms bins / 50+ trials per condition; dPCA 20–30+ neurons / 20+ trials per condition), GPFA latent count by leave-one-neuron-out CV log-likelihood over 2–15 dims, dPCA lambda on held-out trials; the two headline rules carried onto the page — do not z-score firing rates (soft-normalize with a 5 spikes/s constant, or square-root transform) and there is no 90% variance rule, which the skill calls “meaningless for neural data”, requiring parallel analysis or cross-validated reconstruction error instead), parameter-recovery-checker (parameter-recovery-checker.md — the catalog’s first model-identifiability surface; nine-stage simulate-and-refit protocol over 100–1,000+ ground-truth sets with 5–10 optimizer starts each, recovery bands as a table (Pearson r >0.9 good / >0.8 acceptable / <0.7 concerning; bias <10% of range; Bayesian coverage ~95% with <80% concerning; model-recovery confusion diagonal >90% with <70% concerning), the **

  r

  > 0.5 cross-parameter tradeoff cutoff** that fires even when marginal recovery looks fine, the >20% cross-selection rule for calling two models indistinguishable, and the explicit scope exclusion that the skill tests identifiability only, not validity — its own seventh pitfall), lesion-symptom-mapping-guide (lesion-symptom-mapping-guide.md, tagged Neuroscience + Translational Medicine — no prior lesion-mapping coverage; N-gated method routing (VLSM at N ≥ 50 continuous outcome, SVR-LSM at N ≥ 80–100 for distributed patterns, ROI/descriptive only below N = 50, disconnection analysis for white-matter questions, lesion network mapping for network questions), ≥10%-of-sample (or N ≥ 10 patients) per-voxel overlap floor with ≥15% recommended, permutation FWE at 5,000+ permutations thresholded at the 5th percentile of the max-statistic distribution, cost-function masking mandatory at registration with enantiomorphic normalization for large lesions, lesion volume a mandatory covariate; the load-bearing correction argument — spatially correlated lesion voxels violate the independence assumptions behind FDR and parametric corrections, so Bonferroni is “too conservative; almost never detects effects” and uncorrected is never publishable; MATLAB-heavy toolchain named (NiiStat, VLSM2, SVR-LSM toolbox) alongside R LESYMAP, BCBToolkit, ANTs/FSL, no versions pinned upstream and the page says so), netneurotools-guide (netneurotools-guide.md — first netneurotools surface; the eight submodules as a pipeline (datasets → networks → metrics → spatial → modularity → stats → nulls → plotting), pip install "netneurotools[pyvista]" with core deps numpy ≥1.16 / scipy ≥1.4.0 / nibabel ≥3.0.0 / nilearn / bctpy / neuromaps; three parameter rules carried onto the page — consensus modularity needs A[A < 0] = 0 before Louvain, communication metrics need the D = -np.log(W / (np.max(W) + 1)) weight-to-distance transform, and a permutation p-value cannot go below 1 / (n_perm + 1) so the permutation count must be reported with it; framed against generic graph libraries by its spatial-autocorrelation-preserving null models, which is the whole point), pycortex-guide (pycortex-guide.md — first cortical-flatmap/surface-rendering surface; FreeSurfer/fMRIPrep → cortex database → Volume/Vertex/Dataset → 2D flatmap or 3D WebGL viewer → surface geometry, with the shape-must-match-the-transform rule, explicit vmin/vmax for cross-subject comparability, and pack=True for portable datasets). Install-path facts verified this run and carried onto the pages: three of the five (neural-population-analysis-guide, parameter-recovery-checker, lesion-symptom-mapping-guide) declare dependencies: required: [research-literacy], so their single-skill cp -r path copies both directories; netneurotools-guide and pycortex-guide declare no dependency block (they are the collection’s “Claude (AI-assisted)” authored subset rather than “Claude Code Agent”), so their copy path is a single directory and the pages say so explicitly. Underlying-library licenses read from the GitHub license API this run: netneurotools is BSD-3-Clause, pycortex is BSD-2-Clause — both stated in the Pricing row separately from the skill’s own MIT. Two pycortex platform limitations stated inline because they are install-blocking: Linux/macOS only (no Windows path) and the WebGL viewer blocks execution in a plain script (use IPython/Jupyter or autoclose), plus the setuptools wheel numpy cython-first build prerequisite and Python 3.10+. Every page leads its Notes with the collection-wide bolded AI-generated caveat; the three “Claude Code Agent” skills name their cited papers (Cunningham & Yu 2014 / Yu 2009 / Kobak 2016 / Churchland 2012 / King & Dehaene 2014 / Humphries 2021; Heathcote 2015 / Wilson & Collins 2019 / Wagenmakers 2004 / Navarro 2019; Bates 2003 / Kimberg 2007 / Foulon 2018 / Zhang 2014 / Boes 2015 / Sperber 2020) while the two “Claude (AI-assisted)” ones cite no papers at all — the pages point readers at the upstream netneurotools and pycortex documentation instead. Cross-linked into drift-diffusion-model.md, calcium-imaging-analysis-guide.md, spikeinterface-electrophysiology.md, pymc.md, statsmodels.md, fmriprep-tool.md, nilearn-tool.md, freesurfer-tool.md, bnt.md, combraintf.md. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read — all plugins returned are already catalogued; no new entries.

• Three bioSkills hi-c-analysis/ skills + two single-cell/ skills — five new (added 2026-08-08) — Molecular and Cellular Biology directed pass, drawing from the deferred bioSkills queue (GPTomics/bioSkills, MIT). The gap this closes: the catalog had zero 3D-genome coverage — no cooler, no cooltools, no TAD/loop/compartment surface anywhere — despite Hi-C/Micro-C being routine in gene-regulation work. The repo’s 65-directory root listing was re-fetched this run and two previously-unmined categories were located: hi-c-analysis/ (9 skills: compartment-analysis, contact-pairs, hic-data-io, hic-differential, hic-visualization, hichip-plac-loops, loop-calling, matrix-operations, tad-detection) and epitranscriptomics/ (5 skills: m6a-differential, m6a-peak-calling, m6anet-analysis, merip-preprocessing, modification-visualization — listing captured, none catalogued yet; see Deferred). Catalogued the three Hi-C feature-calling layers as a mutually cross-linked set: tad-detection (tad-detection.md, cooler ≥0.10 / cooltools ≥0.7 / bioframe ≥0.7 / HiCExplorer ≥3.7; cooltools.insulation() diamond-window log2 insulation, valley-prominence boundary_strength_{W} + Li/Otsu is_boundary_{W}, the multi-scale [3×bin, 5×bin, 10×bin, …] window sweep, hicFindTADs alternative; the headline boundary-is-reproducible-but-the-domain-partition-is-not rule and its corollary that cross-condition work compares differential score not differential partition, plus insulation-vs-compartment orthogonality), loop-calling (loop-calling.md, cooltools dots HiCCUPS-style 4-background local enrichment with lambda-chunked FDR, chromosight ≥1.6 template correlation, Mustache ≥1.3 scale-space; the depth fork — de-novo needs ~5–10 kb resolution = hundreds of millions to billions of valid pairs, otherwise run APA via cooltools.pileup() on known CTCF/cohesin anchors — consensus-across-callers + convergent-CTCF validation, and differential loops via union anchors + chromosight quantify rather than intersecting two independently-called lists; explicit scoping-out of HiChIP/PLAC-seq/PCHi-C to FitHiChIP/MAPS, which is the sibling hichip-plac-loops skill), compartment-analysis (compartment-analysis.md, cooltools.eigs_cis() on the distance-normalized Pearson-correlated cis matrix at 100 kb–1 Mb (250 kb typical), per-chromosome-arm view_df via bioframe.make_chromarms() to kill the centromere gradient, sort_metric='pearsonr' over eigenvalue ordering, and GC/gene-density phasing so the A sign is not arbitrary). Three shared install-path facts carried onto all three pages: the cooler must be balanced (an unbalanced matrix returns all-NaN insulation, not an error), .mcool files need a single-resolution URI (file.mcool::/resolutions/10000) and never the bare path, and cooltools shifted its API at 0.5→0.7 to view_df/viewframe so >=0.7 must be pinned. Mustache is marked **Unverified —** (the skill names 1.3+ but gives no install command and the PyPI distribution name was not confirmed); HiCExplorer is presented as optional via bioconda with a note about its dependency tree. Also catalogued the two single-cell/ skills flagged as the clean next-run adds on 2026-08-01: hashing-demultiplexing (hashing-demultiplexing.md, Seurat 5.0+ HTODemux/MULTIseqDemux, scanpy 1.10+ hashsolo, pegasus 1.8+ demuxEM, demuxmix 1.4+, GMM-Demux; CLR margin=1-vs-margin=2-for-HTO, positive.quantile = 0.99, demuxEM min_signal = 10.0, hashsolo priors (0.01, 0.8, 0.19) with the doublet prior tracking ~0.8% per 1,000 cells on 10x; method routing clean-HTO→HTODemux, lipid-tags→MULTIseqDemux autoThresh=TRUE, weak-staining/nucleus→demuxmix or demuxEM, few-tags-many-negatives→hashsolo; the two diagnostics — a huge Negative pile means weak staining or ambient, and a near-zero cross-sample doublet rate alongside many expression-flagged doublets means the thresholds are too loose — plus the scoping fact that hashing cannot see same-sample doublets, which is why the page routes to doublet-detection.html), and metabolite-communication (metabolite-communication.md, MEBOCOST 1.0+ enzyme-to-sensor scoring with cutoff_prop = 0.15, n_shuffle = 1000, min_cell_number = 10, permutation_test_fdr < 0.05, transporter calls downgraded; scFEA / Compass / NeuronChat routing with the sharp note that Compass measures reaction favourability, not secretion and so does not license a communication claim, and NeuronChat is neural-only; the headline double-inference warning — enzyme mRNA → protein → activity → flux → metabolite pool → secretion → sensing, none of it measured — requiring metabolomics/MSI/isotope tracing/perturbation for validation, plus the three failure modes: bidirectional “sensor” transporters, ambient RNA, and the human-DB-on-mouse-data silent near-empty result). MEBOCOST install correction: no PyPI distribution was confirmed, so the page installs from git+https://github.com/zhengrongbin/MEBOCOST.git and flags the species-matched config/database path that create_obj() requires. All five tagged Molecular and Cellular Biology only. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read — 19 plugins returned, all already catalogued; no new entries.
• Three bioSkills rna-structure/ skills + geometric-analysis — four new (added 2026-08-08) — Integrative Structural and Computational Biology directed pass. The gap this closes: the entire rna-structure/ category of GPTomics/bioSkills (MIT) was previously unmined, and the catalog’s only RNA-structure surfaces were viennarna-structure-prediction.md (thermodynamic folding) and rfam.md (family lookup) — nothing on evolutionary validation, homolog search with covariance models, or experimental chemical probing. Full 4-skill listing confirmed this run (covariation-analysis, ncrna-search, secondary-structure-prediction, structure-probing); the three with no existing coverage were catalogued as a mutually cross-linked evidence triangle (evolutionary / computational / experimental), and secondary-structure-prediction was deliberately not catalogued because it overlaps viennarna-structure-prediction.md — fold the bioSkills path into that page instead. covariation-analysis (covariation-analysis.md, R-scape 2.0+; -s structure testing against a phylogeny-aware null with proposed and alternative pairs scored separately, --cacofold de novo consensus, E-value target 0.05, the load-bearing ~10% mean alignment power line between “cannot infer” and “rejects the structure”, ~60% average pairwise identity diversity target, independent-substitutions-not-sequence-count rule; outputs .cov/.power/.sorted.cov + R2R diagram; the HOTAIR/Xist/SRA lncRNA negatives as the worked example of an adequately-powered rejection; two rules carried onto the page — a low-power negative says nothing, and only covariation above the null counts, plus the conserved-vs-real boundary). ncrna-search (ncrna-search.md, Infernal 1.1.4+; cmscan/cmsearch/cmalign/cmbuild/cmcalibrate/cmpress/cmfetch, clan de-overlap via --fmt 2 --clanin + grep -v ' = ', GA/TC/NC cutoff semantics with --cut_ga preferred for Rfam, the E-value-scales-with-database-size-Z reproducibility trap and bit-scores-for-custom-CMs corollary, Rfam.cm ships pre-calibrated so cmpress yes / cmcalibrate no, the CM-buys-nothing-on-structure-free-alignments scoping rule that collapses a CM to an HMM; specialized routes to tRNAscan-SE 2.0 / barrnap / RNAmmer / snoscan / snoReport 2.0 / miRDeep2). structure-probing (structure-probing.md, ShapeMapper2 2.1.5+ / ViennaRNA 2.6+ / SEISMIC-RNA 0.20+; SHAPE 1M7/NAI/2A3 four-base vs DMS A/C-only with G/U masked to −999, MaP-vs-RT-stop pipeline split, the three-library MODIFIED/UNTREATED/DENATURED design, per-transcript 2–8% box-plot normalization, Deigan m=1.8 b=−0.6 vs Zarringhalam soft pseudo-energy restraints, SEISMIC-RNA/DREEM per-read clustering for multiple conformations; QC at effective depth ≥5,000, untreated <0.5%, modified ~1–10%, --max-bg 0.05, failing positions reported as −999 not zero; headline rules that reactivity probes constraint, not pairing and that raw reactivities are not comparable across transcripts). geometric-analysis (geometric-analysis.md, structural-biology category, Bio.PDB 1.83+ / NumPy / mkdssp v4; distances and distance matrices, phi/psi/chi dihedrals, Kabsch/SVD Superimposer + QCPSuperimposer, center of mass, radius of gyration, SASA with the probe radius always reported; RSA < 0.20 burial heuristic, ~2.0 Å rigid-core RMSD, TM-score > 0.5 same-fold, Ramachandran outliers read as refinement errors; the headline RMSD-is-not-a-structure-property warning — it depends on superposition and atom selection, and global all-atom RMSD is dominated by the worst-fitting atoms — plus the Superimposer-does-not-solve-correspondence trap and the hetflag id[0] == ' ' water/hetatm filter). Three install-path findings carried onto the pages: (1) bioconda::rscape is 2.0.4.a and bioconda::infernal is 1.1.5 (both checked this run), so both prerequisites have a clean one-line conda install rather than a source build; (2) bioconda::shapemapper is version 1.2 — ShapeMapper 1, not ShapeMapper2 (bioconda/shapemapper2 404s), a real trap the page warns against explicitly, pointing readers at the Weeks-UNC/shapemapper2 release instead with the exact build steps marked **Unverified —**; (3) ShapeMapper2 is Linux-only (Docker/Singularity on macOS), stated inline. geometric-analysis is scoped on its page as the static-structure measurement layer and routed to mdtraj-trajectory-analysis.md / mdanalysis-trajectory.md for the trajectory case and foldseek-structural-search.html for database-scale fold comparison. Directed-pass negative: biomolecular-design-nexus/trrosettarna_mcp (RNA 3D structure prediction MCP, surfaced via a LobeHub listing) 404s on the GitHub API — private or moved, so unverifiable and not catalogued; see Deferred. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read — all plugins re-confirmed present and already catalogued; no new entries.
• Five bioSkills flow-cytometry/ skills — five new (added 2026-08-08) — Immunology and Microbiology directed pass, drawing from the deferred bioSkills queue (GPTomics/bioSkills; MIT via GitHub API, 1,135 stars, pushed 2026-07-25, default branch main). The gap this closes: before this run the entire catalog had exactly one flow-cytometry surface — flowio.md (K-Dense, FCS parsing in Python) — and no coverage at all of compensation, gating, high-dimensional clustering, or cytometry differential testing, despite immunophenotyping being a core Immunology workflow. The flow-cytometry/ directory listing was confirmed this run (8 skills: bead-normalization, clustering-phenotyping, compensation-transformation, cytometry-qc, differential-analysis, doublet-detection, fcs-handling, gating-analysis) and the five with no existing coverage were catalogued as a coherent, bidirectionally cross-linked pipeline (compensation → QC → gating or clustering → differential): compensation-transformation (compensation-transformation.md, flowCore 2.14+ / flowStats 4.14+ / flowWorkspace 4.14+ / CATALYST 1.26+; the load-bearing compensate-then-transform ordering rule — compensation is linear and mathematically invalid after a nonlinear transform, and estimateLogicle() must see already-compensated data so w/a reflect post-compensation negative spread; arcsinh cofactor 5 for CyTOF per Nowicka 2017 F1000Res 6:748 vs ~150 for fluorescence with per-channel flowVS preferred; control brightness ≥ sample brightness per Roederer 2001 Cytometry 45:194; spreading error ∝ √(signal intensity) per Nguyen 2013 Cytometry A 83:306, framed as a panel-design bound not a compensation problem; spectral instruments (Aurora, ID7000) unmix, they do not compensate; CyTOF metal spillover only 1–4% but oxide/impurity real. Two API traps carried onto the page: estimateLogicle() lives in flowWorkspace not flowCore, and flowCore::spillover() returns a list needing [[1]] while flowStats::spillover() returns the matrix directly), cytometry-qc (cytometry-qc.md, flowAI 1.32+ / PeacoQC 1.12+ / flowCore / flowDensity / CATALYST; order-critical margin-removal-before-density rule, PeacoQC MAD 6 and IT_limit 0.55 defaults, flowClean ~30,000-event floor below which CLR frequency tracking under-detects, CyTOF Event_length 10–75, dead cells 10–30% reported as a sample-handling flag not auto-excluded, daily CyTOF retuning for cone fouling; the master-QC-axis insight that a missing or mis-scaled $TIMESTEP silently breaks every time-based tool), gating-analysis (gating-analysis.md, flowWorkspace 4.14+ / openCyto 2.14+ / flowDensity 1.36+ / CytoML 2.14+; hierarchical GatingSet + gs_pop_add() + recompute() + gs_pop_get_stats(), openCyto CSV templates mindensity/tailgate/quantileGate/gate_flowclust_2d, FlowJo round-trip via CytoML, ~50–60 events for CV < 15% Poisson floor, ~1e6 cells for 1e-5 sensitivity, 1e-4–1e-5 rare-event floor; the FMO-not-isotype boundary rule and the time→debris→singlets→live→lineage funnel whose reordering bakes artifacts downstream), clustering-phenotyping (clustering-phenotyping.md, CATALYST 1.26+ / FlowSOM 2.10+ / Rphenograph GitHub-only, returns a list not a vector; prepData() type-vs-state panel annotation, 10×10 over-provisioned SOM grid, maxK 20, 2,000-cell embedding subsample, PhenoGraph k 30; three non-negotiables — never cluster on state markers (activation/phospho/Ki-67 are tested within clusters), embeddings are display-only (never gate or measure distance on UMAP/tSNE), and over-provision-then-metacluster because metaclustering can merge over-fine nodes but cannot split a merged one; set the seed, FlowSOM/Louvain are stochastic), differential-analysis (cytometry-differential-analysis.md, diffcyt 1.22+ / CATALYST / edgeR 4.0+ / limma 3.58+; DA on per-sample-per-cluster counts via edgeR/voom/GLMM vs DS on per-sample-per-cluster arcsinh-medians via limma/LMM, ≥2–3 biological replicates mandatory for a valid error term, BH FDR across clusters and cluster×marker, compositional re-validation with sccomp/scCODA/DCATS when a dominant population shifts, cydar/CITRUS as non-cluster alternatives; the headline warning that per-cell Wilcoxon/t-tests are pseudoreplication that manufactures significance, plus model-batch-in-design-don’t-normalize-it-out). Slug decision: the upstream directory is flow-cytometry/differential-analysis but the catalog already carries the metagenomics differential-abundance.md, so the page is titled Cytometry Differential Analysis at cytometry-differential-analysis.md and states the upstream directory name explicitly in both How to install and Notes so the cp -r path stays copy-pasteable. Rphenograph install-path correction: it is neither on CRAN nor Bioconductor, so the page adds a separate remotes::install_github("JinmiaoChenLab/Rphenograph") line and marks it optional if only FlowSOM is used. All five tagged Immunology and Microbiology + Molecular and Cellular Biology (flow cytometry is core to immunophenotyping and routine in cell biology), all five carry an explicit BiocManager::install(...) prerequisite line, and all five cross-link to flowio.html for the FCS I/O layer. Front-matter names (bio-flow-cytometry-*), package versions, workflow stages, thresholds and cited papers verified against the GitHub SKILL.md files this run. Manifest sweep: anthropics/life-sciences/.claude-plugin/marketplace.json re-read — all plugins re-confirmed present and already catalogued; no new entries.

• ChemGraph — one new (added 2026-08-08) — Chemistry directed pass, resolving the candidate deferred 2026-08-01. Both blockers cleared from primary sources this run: the repo is Apache-2.0 (GitHub license API), and the MCP tool surface was read directly from src/chemgraph/mcp/mcp_tools.py — four @mcp.tool() functions (molecule_name_to_smiles via PubChem, smiles_to_coordinate_file, run_ase, extract_output_json). PyPI chemgraph 0.6.0 (2026-07-19), requires_python >=3.10. The key finding that made a followable-verbatim page possible: src/chemgraph/mcp/server_utils.py shows run_mcp_server() defaults to --transport stdio, with --transport streamable_http / --host / --port (default_port=9003 for mcp_tools) as opt-in — so the page leads with a plain claude mcp add --transport stdio chemgraph -- /path/to/chemgraph-env/bin/python -m chemgraph.mcp.mcp_tools rather than the README’s HTTP invocation, and the HTTP form is presented separately and labelled a long-lived service. Three install-path corrections carried onto chemgraph.md: (1) the registration must use the absolute path to the venv’s python, not bare python, because Claude Code launches the process outside the activated shell; (2) the README’s docker run ghcr.io/argonne-lcf/chemgraph:latest starts JupyterLab, not the MCP server — explicitly called out as not an install path; (3) the /mcp HTTP endpoint path is the FastMCP streamable_http_app() default and is not documented upstream, so it is marked **Unverified —**. Pricing row separates the Apache-2.0 wrapper from the per-engine terms of the backends it drives (TBLite/xTB LGPL-3.0 and NWChem ECL-2.0 free, ORCA free-for-academic-after-registration, Gaussian commercial); only the [calculators] extra (TBLite) works out of the box. Sibling HPC/Parsl MCP modules in the same package (mace_mcp_parsl, graspa_mcp_parsl, xanes_mcp_parsl, ase_mcp_hpc, data_analysis_mcp, hpc_misc_mcp) are named on the page but flagged undocumented/unverified. Safety framing: run_ase executes real computations locally and writes files, so the page tells the reader to use a scratch directory and warns that a casually-requested DFT job burns real wall time. Cross-linked bidirectionally with xtb-mcp-server.md (input prep vs execution) — that page’s stale “not yet catalogued pending license and tool-list confirmation” line was replaced with a live link. Manifest sweep: all plugins in anthropics/life-sciences/.claude-plugin/marketplace.json re-confirmed present and already catalogued; no new entries.

• Four ToolUniverse drug-discovery skills — four new (added 2026-08-02) — Drug Repurposing and Discovery directed pass. The two seed queries (drug repurposing MCP server, DrugBank / target prioritization MCP install) returned no uncatalogued, license-clear MCP server: openpharma-org/drugbank-mcp-server is the already-catalogued drugbank.md, Augmented-Nature/ChEMBL-MCP-Server remains deferred as redundant with the first-party chembl.md connector, cyanheads/pubchem-mcp-server duplicates pubchem.md’s data source (fold in as an alt install path only after an audit), and FDB MedProof MCP (GA 2026-04, medication decision support) is an enterprise product with no published enable path. The pass therefore drew from the ToolUniverse skills queue (mims-harvard/ToolUniverse, Apache-2.0, ~68 skills; skills/ directory listing re-confirmed this run), cataloguing the four strongest DRD skills with no existing coverage — all four verified against their GitHub SKILL.md front-matter and body this run, and all four carry disable-model-invocation: true upstream (so each page tells the reader to invoke the skill explicitly). tooluniverse-admet-prediction (tooluniverse-admet-prediction.md, Chemistry + DRD) — five-phase ADMET profiling: PubChem identity resolution → ADMET-AI + SwissADME physicochemical/Lipinski/TPSA/logP/PAINS → BBB, oral bioavailability, CYP, clearance, distribution → AMES/DILI/hERG/carcinogenicity predictions cross-checked against experimental PubChemTox (LD50/LC50, GHS, target organs) and nuclear-receptor/stress-response panels → a 13-category pass/warn/fail scorecard with T1–T3 evidence tiers; upstream rules state experimental data outranks prediction and a hERG/AMES/DILI FAIL is program-limiting until wet-lab evidence refutes it. Cross-linked to admetlab-mcp.md (self-hosted ADMETlab 3.0 alternative), inductive-bio.md (enterprise), and tooluniverse-small-molecule-discovery.md. tooluniverse-dose-response (tooluniverse-dose-response.md, Chemistry + DRD + MCB) — 4PL/Hill curve fitting via DoseResponse_calculate_ic50 and DoseResponse_compare_potency: linear-not-log concentration scale, zero concentrations dropped, ≥4 points spanning both plateaus, Hill slope >1.5 or <0.5 flagged, r² ≥ 0.95 required before a potency comparison is endorsed and r² < 0.90 forces manual inspection, biphasic/non-monotonic data rejected from 4PL outright, units always reported. Upstream scope exclusions carried onto the page (image-derived dose-response and survival/regression route to sibling skills). Cross-linked to tooluniverse-drug-synergy.md, statsmodels.md, scikit-learn.md. tooluniverse-cell-line-profiling (tooluniverse-cell-line-profiling.md, DRD + MCB + Translational Medicine) — cancer cell-line selection across Cellosaurus (STR/contamination flags), DepMap (tissue, cancer type, MSI), COSMIC + cBioPortal CCLE mutations, HPA expression (only 10 lines supported), CellMarker, PharmacoDB IC50/AAC over GDSC/CCLE/PRISM, SYNERGxDB ZIP synergy, ranked by a weighted score out of 27 (mutation match ×3, co-mutation simplicity ×2, gene dependency ×2, drug data ×1, practical ×1); DepMap Chronos < −0.5 is the essentiality cut-off. Upstream limitation carried onto the page: the DepMap API returns metadata only, so per-cell-line Chronos scores need the bundled depmap_gene_dependency.py script or the depmap.org portal — not a live tool call. Cross-linked to depmap.md, cbioportal.md, cosmic-database.md, tooluniverse-drug-synergy.md, tooluniverse-drug-target-validation.md. tooluniverse-chemical-sourcing (tooluniverse-chemical-sourcing.md, Chemistry + DRD) — 10 tools (PubChem_get_CID_by_compound_name, PubChem_get_compound_properties_by_CID, ChEMBL_get_molecule, ZINC_search_compounds, ZINC_get_compound, Enamine_search_catalog, Enamine_get_compound, eMolecules_search, eMolecules_get_compound, Mcule_get_compound) over ZINC 230M+ / Enamine ~4M in-stock / eMolecules 8M+ / Mcule 40M+; Tanimoto ≥ 0.7 analog fallback, in-stock-beats-price (make-on-demand 2–4 weeks), purity floors ≥95% screening / ≥98% dose-response / ≥99% reference standard, >5× median price treated as a salt-form/purity signal with per-mg normalization, SMILES searches preferred over names. Distinguished on the page from the one-step commercial-availability phase of tooluniverse-small-molecule-discovery.md; cross-linked to zinc-database.md, pubchem.md, chembl.md, sar-analysis.md. Manifest sweep: all 21 plugins in anthropics/life-sciences/.claude-plugin/marketplace.json re-confirmed present and already catalogued; no new entries.
• Two OMOP MCP servers + three bioSkills clinical-biostatistics skills — five new (added 2026-08-02) — Translational Medicine directed pass. The FHIR seed query returned nothing new (the catalog already carries fhir-wso2.md, fhir-momentum.md, healthlake-mcp.md, fhir-developer.md), so the pass pivoted to the two uncovered TM surfaces: OMOP/OHDSI and clinical-trial biostatistics. OMOPHub MCP Server (omophub-mcp.md; OMOPHub/omophub-mcp, MIT via GitHub API, npm @omophub/omophub-mcp 1.5.3, engines.node >=20, repo pushed 2026-07-28) — 11 tools over the OHDSI standardized vocabularies (search_concepts, get_concept, get_concept_by_code, semantic_search, find_similar_concepts, map_concept, get_hierarchy, explore_concept, fhir_resolve, fhir_resolve_codeable_concept, list_vocabularies) plus resources omophub://vocabularies[/{id}] and guided prompts phenotype-concept-set / code-lookup; hosted API at api.omophub.com needs a free oh_-prefixed key from dashboard.omophub.com/api-keys; free-tier limits/prices not published → Pricing row marked **Unverified —**. Install-path correction caught this run: the upstream README’s claude mcp add omophub -- npx -y @omophub/omophub-mcp omits the key, so the page uses the --env OMOPHUB_API_KEY=… form and says why; the Docker/HTTP paths are flagged as long-lived services, and the README’s own caveat that Claude Desktop’s Custom Connectors UI is OAuth-only (cannot send the Bearer header the hosted mcp.omophub.com endpoint needs) is carried onto the page. pyomop (pyomop.md; dermatologist/pyomop, GPL-3.0 via GitHub API — copyleft, called out; PyPI 6.4.0, requires_python >=3.11,<4.0, 64 stars, pushed 2026-07-28, default branch develop not main) — a Python OMOP CDM toolkit that ships an MCP server with 8 tools (create_cdm, create_eunomia, get_usable_table_names, get_table_columns, get_single_table_info, run_sql, check_sql, example_query) over SQLAlchemy against SQLite/PostgreSQL/MySQL; console scripts pyomop-mcp-server (stdio) and pyomop-mcp-server-http, extras [llm] and [http]. Two install-path corrections: the README’s Claude Desktop snippet uses a top-level "servers" key (VS Code form, not Claude Desktop’s "mcpServers") and depends on uv run, so the page shows the mcpServers + console-script form; and the page states the stdio verify step is a one-shot Ctrl-C check. Safety framing on the page: create_cdm/create_eunomia are upstream-restricted to local SQLite but run_sql is not, so point it at a real CDM only through a read-only role. PyPI leaves license_expression empty → GPL-3.0 marked as read from the GitHub license field. The two are cross-linked as vocabulary-service vs CDM-query-server, and omophub-mcp is cross-linked to medical-terminologies-mcp.md. Three bioSkills clinical-biostatistics/ skills (GPTomics/bioSkills, MIT — 12-skill directory listing confirmed this run: adaptive-designs, bayesian-trials, categorical-tests, cdisc-data-handling, effect-measures, logistic-regression, missing-data-sensitivity, multiplicity-graphical, power-and-sample-size, subgroup-analysis, survival-analysis, trial-reporting; the whole category was previously unmined): adaptive-designs (adaptive-designs.md, rpact 4.2+ / gsDesign 3.6+ / gsDesign2 1.1+ / adaptr / simtrial / BOIN / dfcrm / trialr / escalation; O’Brien-Fleming / Pocock / Lan-DeMets spending, blinded Friede-Kieser vs unblinded Cui-Hung-Wang vs Mehta-Pocock promising zone, seamless Phase 2/3 arm selection, enrichment, RAR, IDMC firewall, bias-corrected final estimates; anchored to FDA 2019 Final Adaptive Designs Guidance, FDA 2022 Master Protocols, and the ICH E20 Step 2b draft of 2025-06-25 which the skill explicitly says is NOT final), missing-data-sensitivity (missing-data-sensitivity.md, upstream name bio-clinical-biostatistics-missing-data; ICH E9(R1) estimand-first, Little’s MCAR test on the CDISC DS domain, MMRM with UN covariance + Kenward-Roger via R mmrm ≥0.3, reference-based MI J2R/CR/CIR/LMCF via rbmi ≥1.5 per Carpenter-Roger 2013, Permutt delta-adjustment / tipping point in residual-SD units, pattern-mixture CCMV/NCMV/ACMV, the Cro-vs-Bartlett Rubin’s-vs-frequentist variance reconciliation, von Hippel’s m ≥ 100 × FMI imputation-count rule; the statsmodels.mixedlm-has-no-Kenward-Roger trap stated explicitly), trial-reporting (trial-reporting.md, tableone 0.9+ primary with rpy2 bridging to R mmrm/gMCP 0.8+/RBesT 1.6+; CONSORT 2025 + SPIRIT 2025 + ICH E9(R1) + FDA 2023 covariate adjustment, ITT/FAS/PP/Safety population definitions, Table 1 with SMD > 0.1 imbalance flag and no baseline significance testing, Bretz-Maurer graphical multiplicity, CONSORT flow diagram + Item 21c + Item 4). All three cross-linked to each other and to clinical-trial-protocol.md / scikit-survival.md. Front-matter, package/version tables, thresholds and cited guidances verified against the GitHub SKILL.md files this run. Manifest sweep: all 21 plugins in anthropics/life-sciences/.claude-plugin/marketplace.json re-confirmed present and already catalogued; no new entries.
• Five awesome_cognitive_and_neuroscience_skills skills — five new (added 2026-08-02) — Neuroscience directed pass, resolving the collection deferred 2026-07-12 pending an audit. Audit result: repo is MIT (SPDX confirmed via the GitHub API; default branch is master, not main — main raw URLs 404), 34 stars, last push 2026-07-13, and it ships a real Claude Code plugin marketplace (/plugin marketplace add HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills + /plugin install awesome-cognitive-and-neuroscience-skills@awesome-cognitive-and-neuroscience-skills; marketplace name and plugin name are identical, source: "./"). skills/ holds 50 directories, ~10 of which are meta/contribution skills (contribute-skill, contribute-skills-via-pr, verify-skill, paper-to-skill, repo-to-skill, share-case, share-usage, research-literacy, markdown-report-writing) — hence the marketplace’s “40 skills” count vs the README’s “50”. Every skill carries review_status: ai-generated and the README states the content “has not been individually verified by human domain experts”; each catalogued page therefore leads its Notes with a bolded AI-generated caveat naming the skill’s cited papers. Catalogued the five with no existing catalog coverage: deeplabcut (deeplabcut.md, markerless pose estimation — 9-stage pipeline, maxiters 100,000 / saveiters 5,000 / batch 8→4→2 on GPU pressure, ResNet-50/101 + MobileNet v2 + EfficientNet-b0, SuperAnimal zero-shot scale list [200,300,400], multi-animal via fasterrcnn_resnet50_fpn_v2, pip install "deeplabcut[gui]" + separately-matched PyTorch/CUDA; DeepLabCut itself is LGPL-3.0, not covered by the skill’s MIT), calcium-imaging-analysis-guide (calcium-imaging-analysis-guide.md, rigid-vs-non-rigid motion correction at ≤10% FOV shift, CNMF / CNMF-E / Cellpose ROI extraction, neuropil r ≈ 0.7 subtracted before dF/F, 8th–20th-percentile rolling baseline over 30–60 s, OASIS/FOOPSI/CASCADE/MLSpike deconvolution framed as rate estimates not spike trains, QC at SNR > 3 and skewness > 0.5; CaImAn is conda-only, Suite2p/Cellpose pip), drift-diffusion-model (drift-diffusion-model.md, 8-stage protocol, RT cutoffs <200 ms and >3,000–5,000 ms, EZ-diffusion / fast-dm / PyDDM / HDDM selection guidance, BIC for MLE vs DIC/WAIC for Bayesian, posterior predictive checks on RT quantiles, parameter recovery as a pipeline test), optogenetics-protocol-designer (optogenetics-protocol-designer.md, manipulation-type → opsin → light → pulse → fiber → controls; ChR2(H134R)/ChETA/Chronos/CheRiff/ChrimsonR/ChRmine/C1V1(TT) vs eNpHR3.0/eArch3.0/stGtACR2/SwiChR++ vs bistable SSFO/SOUL/SwiChR; irradiance specified at target depth not fiber tip, duty cycle inside channel kinetics, opsin-negative + off-spectrum + within-subject light-off controls; Read-only, no software executed), signal-detection-analysis (signal-detection-analysis.md, d′ = z(H) − z(FA), c = −0.5[z(H)+z(FA)], ln β = d′·c, c′ = c/d′, log-linear +0.5 extreme-value correction instead of clipping, d_a / A_z when the ROC slope ≠ 1). Install-path detail caught in the audit: four of the five declare dependencies: required: [research-literacy] in front-matter, so the single-skill cp -r path must copy skills/research-literacy as well — every page does. Front-matter, workflow stages, thresholds and cited papers verified against the GitHub SKILL.md files this run. Manifest sweep: all 21 plugins in anthropics/life-sciences/.claude-plugin/marketplace.json re-confirmed present and already catalogued (positions 19–21 are consensus, cortellis, adisinsight — all covered); no new entries.
• Five bioSkills single-cell/ skills — five new (added 2026-08-01) — Molecular and Cellular Biology directed pass, drawing from the deferred bioSkills queue (GPTomics/bioSkills, MIT). Full single-cell/ directory listing confirmed this run (17 skills: batch-integration, cell-annotation, cell-communication, clustering, cnv-inference, data-io, differential-abundance, doublet-detection, hashing-demultiplexing, lineage-tracing, markers-annotation, metabolite-communication, multimodal-integration, perturb-seq, preprocessing, scatac-analysis, trajectory-inference). Catalogued the five with no existing catalog coverage: doublet-detection (doublet-detection.md, scDblFinder 1.16+ / Scrublet via scanpy 1.10+ / DoubletFinder post-Nov-2023 *_v3-removed; per-sample-before-integration rule, expected rate dbr.per1k = 0.008 ≈ 0.8% per 1,000 recovered cells, total-lane not demultiplexed-subset rate for multiplexed pools, modelHomotypic() adjustment, non-bimodal Scrublet histogram tuning, flag-and-inspect over blind deletion; upstream tool_type: mixed), cnv-inference (cnv-inference.md, inferCNV 1.18+ reference-based / copyKAT 1.1+ + SCEVAN 1.0+ reference-free / Numbat 1.4+ allele-aware; gene-order file, sex-matched in-sample normal reference, HMM i6, cutoff 0.1 droplet vs 1 Smart-seq, lineage-marker + allele validation before calling a cell malignant; tagged MCB + Translational Medicine), perturb-seq (perturb-seq.md, Pertpy 0.9+ / SCEPTRE 0.10+ / Mixscape / scMAGeCK; mixture-model guide posterior over flat thresholds, escaper removal, replication unit is the transfection not the cell → pseudobulk ≥2–3 replicates, E-distance in a fixed PCA embedding, Milo/scCODA compositional split, whole-perturbation holdout vs additive baseline for foundation models, low-MOI discards 70–90% of cells), scatac-analysis (scatac-analysis.md, Signac 1.13+ / Seurat 5.0+ default, ArchR 1.0+, SnapATAC2 2.x; TSS + nucleosome QC, TF-IDF→LSI with the depth-correlated component 1 diagnosed and dropped, per-cluster consensus peak merge, depth-covariate logistic-regression DA, chromVAR against GC-matched backgrounds, TOBIAS footprinting, AMULET/scDblFinder homotypic-vs-heterotypic doublets, binarization decision), lineage-tracing (lineage-tracing.md, Cassiopeia 2.0+ / Startle / CoSpar 0.3+; CRISPR scars vs LARRY/CellTag static barcodes vs mtDNA, character-matrix construction, VanillaGreedy/ILP/Hybrid/NeighborJoining solvers, Robinson–Foulds + triplets-correct cross-solver robustness, ~10 UMI/cell floor, drop >~50% missing characters, informative-character rule, library complexity ≫ founders). Front-matter, tool/version tables, workflow stages and thresholds verified against the GitHub SKILL.md files this run. Install-path correction caught and fixed: lineage-tracing cannot use pip install cassiopeia-lineage — the PyPI distribution is still at 1.0.4 (checked this run) while the skill targets 2.0+, so the page installs Cassiopeia from git+https://github.com/YosefLab/Cassiopeia@master and says why; cospar on PyPI is at 0.5.0 so >=0.3 resolves fine. Cross-linked into single-cell-rna-qc.md, scanpy.md, anndata.md, celltypist-cell-annotation.md, popv-cell-annotation.md, cbioportal.md, cosmic-database.md, sgrna-design-guide.md, macs3-peak-calling.md, homer-motif-analysis.md, jaspar-database.md, deeptools.md, scvelo.md, cellrank-mcp.md. Manifest sweep: all 18 plugins currently listed in anthropics/life-sciences/.claude-plugin/marketplace.json re-confirmed present and already catalogued (no new entries; biorxiv and clinical-trials remain flagged DOA).
• Rosetta MCP Server + four bioSkills structural-biology skills — five new (added 2026-08-01) — Integrative Structural and Computational Biology directed pass. Rosetta MCP Server (Arielbs/rosetta-mcp-server, npm rosetta-mcp-server 1.3.1, MIT per package.json only) — 18 tools over Rosetta / PyRosetta / Biotite: run_rosetta_scripts, pyrosetta_score (with per_residue), validate_xml (+ XSD schema check), rosetta_scripts_schema, xml_to_pyrosetta, rosetta_to_biotite / biotite_to_rosetta / translate_rosetta_script_to_biotite (~21 mappings; design ops flagged Rosetta-only), get_rosetta_help (movers/concepts/score functions, live docs), pyrosetta_introspect, search_rosetta_web_docs, get_rosetta_web_doc, get_cached_docs, get_rosetta_info, python_env_info, check_pyrosetta, install_pyrosetta_installer, find_rosetta_scripts. Node 14+ / Python 3.8+; npm install -g + a uv venv holding pyrosetta-installer and biotite; stdio, configured entirely through PYTHON_BIN + ROSETTA_BIN. Upstream README documents Cursor only — the Claude Code claude mcp add --transport stdio … --env … and Claude Desktop JSON snippets on the page are constructed equivalents, called out as such. Written up by RosettaCommons 2025-10-20. Catalogued rosetta-mcp-server.md, tagged Integrative Structural and Computational Biology. Rosetta/PyRosetta themselves need a University of Washington licence (free academic) — stated in the Pricing row; the Biotite translation tools exist partly as the licence-free analysis path. Four bioSkills structural-biology/ skills (GPTomics/bioSkills, MIT — 10-skill directory listing confirmed this run): structure-validation (structure-validation.md, resolution / R-free-minus-R-work overfitting gap / within-structure B-factor MAD z-scoring / MolProbity clashscore-Ramachandran-rotamer-cis-nonPro via phenix.molprobity / pLDDT+PAE bands and phenix.process_predicted_model trimming / cryo-EM FSC 0.143 half-map vs 0.5 map-model / NMR ensemble spread), binding-site-detection (binding-site-detection.md, fpocket 4.1+ alpha-spheres, P2Rank 2.4+, CASTp, DoGSiteScorer, mdpocket over MD ensembles; the “geometric cavity is a hypothesis” and “druggability scores were trained on holo sets so they under-detect apo/shallow/cryptic” rules; tool_type: mixed — two of the five are external web servers, flagged for data-sharing policy), interface-analysis (interface-analysis.md, cutoff-with-rationale 4–5 Å heavy-atom vs 8 Å Cα vs 3.5–4.0 Å H-bond, NeighborSearch contacts, ShrakeRupley BSA = SASA(A)+SASA(B)−SASA(complex) halved per partner, PDBePISA CSS crystal-packing check, biological-assembly-not-asymmetric-unit rule, epitope/ligand footprints), structure-preparation (structure-preparation.md, PDBFixer 1.9+ / OpenMM 8.1+ / reduce-Reduce2 / PROPKA 3 / PDB2PQR 3.0 — His HID/HIE/HIP tautomers, Asn/Gln/His flips, pKa-shifted Cys/Lys/Asp/Glu at a stated pH, SEQRES-vs-modelled gap finding, MSE/PTR replacement, short loops as disorder hypotheses, PQR for Poisson–Boltzmann). Install: git clone https://github.com/GPTomics/bioSkills + ./install-claude.sh --categories "structural-biology" or single-skill cp -r bioSkills/structural-biology/<skill> ~/.claude/skills/. Front-matter names, tool/version lists, workflow stages and external deps verified against the GitHub SKILL.md files this run. Cross-linked into pdb.md/pdbe.md/alphafold.md/autodock-vina-docking.md/smina-molecular-docking.md/diffdock.md/openmm-mcp.md/gromacs-mcp.md/epitope-prediction.md/complex-portal.md/intact.md. Manifest sweep: all 21 anthropics/life-sciences plugins re-confirmed present and already catalogued (no new entries). Directed-pass negatives reconfirmed: no cryo-EM MCP/Skill for RELION or cryoSPARC exists (the cryosparc-tools Python API is the unwrapped building block; cryo-mcp is a blockchain-data name collision), and no standalone TM-align/Foldseek/FoldMason MCP — TM-align/jFATCAT and Foldseek remain reachable only through the catalogued protein-mcp-server.md.
• bioSkills metagenomics + immunoinformatics + microbiome skills — five new (added 2026-08-01) — surfaced on the Immunology and Microbiology directed pass by drawing from the deferred bioSkills queue (GPTomics/bioSkills, MIT). Verified against per-skill SKILL.md front-matter + the metagenomics/ directory listing this run and catalogued: metaphlan-profiling (metaphlan-profiling.md, MetaPhlAn 4.1+ clade-specific marker profiling to species/SGB via Bowtie2/minimap2, --stat_q quantile-truncated coverage, unknown-fraction rescaling, --index pinning as a batch variable, and the cell-fraction-vs-Kraken-read-fraction rule), functional-profiling (functional-profiling.md, HUMAnN 3 tiered search — MetaPhlAn prescreen → Bowtie2 ChocoPhlAn pangenome → DIAMOND vs UniRef90/50 — with RPK→CPM normalization, UNMAPPED/UNINTEGRATED retention, KO/EC/GO regrouping, and eggNOG/dbCAN/antiSMASH alternatives), strain-tracking (strain-tracking.md, inStrain popANI/conANI ≥99.999% over ≥50% breadth at ≥5×, StrainPhlAn nGD, MIDAS2, StrainGE, metaSNV v2, dRep + skani/fastANI/MASH), immunogenicity-scoring (immunogenicity-scoring.md, NeoFox ~16-feature annotation, PRIME2.0 + MixMHCpred v3.0+, BigMHC-IM, Łuksza/Balachandran fitness model agretopicity + foreignness, pVACtools tiering, TPM ≥1 / RNA VAF ≥0.25 gates), differential-abundance (differential-abundance.md, ALDEx2 / ANCOM-BC2 / MaAsLin2-3 / LinDA / ZicoSeq / LEfSe / q2-composition consensus per the Nearing benchmark, prevalence filter + effect-size floor, why DESeq2/edgeR misfire). All tagged Immunology and Microbiology; immunogenicity-scoring also Translational Medicine. Install: git clone https://github.com/GPTomics/bioSkills + ./install-claude.sh --categories "metagenomics" (or "immunoinformatics" / "microbiome") or single-skill cp -r bioSkills/<cat>/<skill> ~/.claude/skills/. Front-matter names, tool/database lists, workflow stages, thresholds, and external deps verified against the GitHub SKILL.md files this run. Cross-linked to the already-catalogued kraken-classification, amr-detection, amplicon-processing, taxonomy-assignment, neoantigen-prediction, mhc-binding-prediction, epitope-prediction. Note on immunogenicity-scoring: MIT covers the skill text only — MixMHCpred/PRIME carry academic-use terms from their own authors, called out inline. Manifest sweep: all 21 anthropics/life-sciences plugins re-confirmed present and already catalogued (no new entries); note the manifest lives at .claude-plugin/marketplace.json, not repo-root marketplace.json (root path 404s).
• XTB MCP Server + Materials Project MCP Server — two new (added 2026-08-01) — surfaced on the Chemistry directed pass (quantum chemistry MCP server Psi4 xtb ORCA DFT seed query → GitHub) plus a manifest read of the mcp.science collection. XTB MCP Server (PhelanShao/xtb-mcp-server, MIT, clone-install — not on PyPI) — semi-empirical quantum-chemistry input preparation for the Grimme-group xtb code across GFN0/GFN1/GFN2/GFN-FF, 10 tools (generate_xtb_input_package, generate_xcontrol_file, validate_xtb_input_files, convert_structure_file_format, explain_xtb_parameters, generate_enhanced_sampling_input, generate_wavefunction_analysis_input, generate_oniom_input, generate_spectroscopy_input, analyze_trajectory), stdio, Desktop JSON from the README + a constructed claude mcp add equivalent; catalogued xtb-mcp-server.md, tagged Chemistry. Two gaps marked **Unverified —** inline: the README states no minimum Python version and does not say whether the xtb binary must be on PATH separately (the tool names are all generate/validate, so the server itself likely only writes decks). License statement appears only in the README’s Chinese section. Materials Project MCP Server (pathintegral-institute/mcp.science, MIT) — 7 tools over the Materials Project crystal-structure database (search_materials_by_formula, select_material_by_id, get_structure_data CIF/POSCAR, create_structure_from_poscar, plot_structure, build_supercell, moire_homobilayer), free account + MP_API_KEY, launched with the collection-wide uvx mcp-science materials-project; catalogued materials-project-mcp.md, tagged General-Purpose Utilities (adjacent-domain materials science). Licenses, tool lists, transports, and install snippets verified against the GitHub READMEs this run. Manifest sweep: all 21 anthropics/life-sciences plugins re-confirmed present and already catalogued (no new entries; bio-research correctly lives in knowledge-work-plugins, not this marketplace).
• ADMETlab MCP Server + Drug Pipeline MCP Server — two new (added 2026-07-26) — surfaced on the Drug Repurposing and Discovery directed pass (ADMET prediction Claude skill/MCP / drug repurposing agent MCP server seed queries → GitHub). ADMETlab MCP Server (ToxMCP/admetlab-mcp, Apache-2.0, install-from-source — not on PyPI) — self-hostable HTTP MCP wrapping the ADMETlab 3.0 API, 4 tools (wash_molecule, render_molecule_svg, predict_admet, fetch_admet_csv), keyless, Python 3.10+, uvicorn admetlab_mcp.transport.http:app --port 8200 long-lived server + claude mcp add --transport http / Desktop mcp-remote; catalogued admetlab-mcp.md, tagged Chemistry + Drug Repurposing and Discovery. Free/self-hosted counterpart to the enterprise-gated inductive-bio.md ADMET connector. Drug Pipeline MCP Server (DasClown/drug-pipeline-mcp, MIT, PyPI drug-pipeline-mcp) — pharmaceutical R&D pipeline-intelligence aggregator, 6 source-traceable tools (search_trials, get_approvals, get_safety_data, get_drug_label, get_eu_approvals, composite drug_pipeline) over ClinicalTrials.gov + openFDA + RxNorm + PubMed + EMA + DailyMed + Open Targets + MyChem.info, keyless, Python 3.10+, stdio (uvx drug-pipeline-mcp) + optional HTTP; catalogued drug-pipeline-mcp.md, tagged Drug Repurposing and Discovery + Translational Medicine. Distinct from the discrete openFDA/ClinicalTrials.gov entries by its cross-source synthesis + EMA approval coverage (no ML/predictions — every output cites NCT/FDA app #/PMID). Licenses, tool lists, transports, keyless status, and install snippets verified against the GitHub READMEs + PyPI this run. DrugBank MCP Server (drugbank.md) last_verified bumped 2026-07-26 (repo active, MIT, 17,430 drugs, npm/node path unchanged). Open Targets standalone MCP + Inductive Bio ADMET reconfirmed already covered.
• OpenFDA MCP Server (cyanheads) + CMS data.gov MCP — two new (added 2026-07-26) — surfaced on the Translational Medicine directed pass (OpenFDA drug adverse event MCP server / CMS Blue Button Medicare claims MCP server seed queries → GitHub). OpenFDA MCP Server (cyanheads) (cyanheads/openfda-mcp-server, Apache-2.0, npm @cyanheads/openfda-mcp-server) — federates the full openFDA API (drugs/food/devices/veterinary/recalls) with 14 tools including openfda_search_device_clearances (510k/PMA), openfda_search_drug_shortages, openfda_search_recalls, plus opt-in DuckDB SQL; public HTTP instance https://openfda.caseyjhand.com/mcp, bunx/npx stdio, optional OPENFDA_API_KEY; catalogued openfda-mcp-server.md, tagged Drug Repurposing and Discovery + Translational Medicine. Fills the device-clearance gap the deferred Augmented-Nature/OpenFDA-MCP-Server left open (that variant still lacks a copy-pasteable registration snippet + explicit license). Distinct from the drug-only openfda.md (ythalorossy, 7 tools) — cross-linked both ways; openfda.md last_verified bumped 2026-07-26. CMS data.gov MCP Server (clarifyhealth/cms-datagov-mcp-server, MIT, npm @clarify/cms-datagov-mcp-server) — 5 tools + resource templates over data.cms.gov statistical datasets (cms_search_datasets, cms_get_dataset, cms_query_dataset ≤5,000 rows, cms_get_dataset_stats, cms_get_csv_link), Node 18+, keyless, clone + npm run build + npm link; catalogued cms-datagov-mcp.md, tagged Translational Medicine. Distinct from the Anthropic CMS Coverage policy MCP (cms-coverage.md, LCD/NCD) — this serves the broader statistical data catalog. License, tool lists, transports, keyless status, and install snippets verified against the GitHub READMEs + npm this run. openpharma-org/fda-mcp and openpharma-org/medicare-mcp deferred on unspecified upstream license (see Deferred).
• Protein MCP Server + ChimeraX MCP Server — two new (added 2026-07-25) — surfaced on the Integrative Structural and Computational Biology directed pass (RCSB PDB MCP server protein structure ChimeraX MCP / molecular dynamics MCP server GROMACS OpenMM cryo-EM Claude skill seed queries → GitHub). Protein MCP Server (cyanheads/protein-mcp-server, Apache-2.0, npm @cyanheads/protein-mcp-server) — keyless federated structure engine over RCSB PDB + AlphaFold DB + 3D-Beacons + UniProt + InterPro + Foldseek, 7 tools (protein_search_structures, protein_get_structure, protein_find_similar mmseqs2/Foldseek homologs, protein_track_ligands, protein_compare_structures TM-align/jFATCAT, protein_analyze_collection, protein_get_annotations), stdio via npx -y + public HTTP instance https://protein.caseyjhand.com/mcp; catalogued protein-mcp-server.md, tagged Integrative Structural and Computational Biology. ChimeraX MCP Server (mahynotch/chimerax-mcp, MIT, PyPI chimerax-mcp) — natural-language UCSF ChimeraX driver, 39 tools (structure/editing/visualization/measurement/selection), auto-launches ChimeraX, pip install chimerax-mcp + claude mcp add -s user chimerax -- chimerax-mcp, Python 3.10+, requires local ChimeraX (post-2024-03-18); catalogued chimerax-mcp.md, tagged Integrative Structural and Computational Biology. License, tool lists, transports (stdio/HTTP), keyless/prereq status, and install snippets verified against the GitHub READMEs + npm/PyPI this run. Distinct from the RCSB-only pdb.md (Protein MCP federates predicted models + Foldseek/TM-align) and the headless pymol.md (ChimeraX is an interactive GUI driver); all cross-linked.

Flagged for review

• Owkin Pathology Explorer Connector (owkin.md) — flagged 2026-08-15 during the link/pricing recheck (oldest entry in the catalog, last_verified 2026-05-24). owkin.com resolves fine, but the homepage no longer names Pathology Explorer anywhere — the product narrative now leads with K Pro, “the first AI agent connecting research and care”. The owkin entry is still present in anthropics/life-sciences/.claude-plugin/marketplace.json (position 9, re-confirmed this run), so the documented install path stands; the entry was not removed and its last_verified was deliberately not bumped. This may be no more than homepage rotation, but combined with the verifier’s standing verification: degraded (MCP endpoint vendor-gated, so the tool list has never been publicly resolvable) it warrants one targeted check. Next run: fetch the Owkin newsfeed and product pages directly rather than the homepage, and determine whether Pathology Explorer has been folded into K Pro — if so the page needs retitling and rescoping, not just re-verifying. Un-flag once the product’s current status is confirmed either way.

• Rosetta MCP Server (rosetta-mcp-server.md) — flagged 2026-08-01: the license is declared only in package.json / the npm registry metadata (MIT); no LICENSE file is committed to Arielbs/rosetta-mcp-server (top-level listing confirmed this run). Catalogued because the npm manifest is an explicit declaration (unlike ChemCP / PLSDB, where nothing is declared anywhere), but the Pricing row marks the redistribution terms **Unverified —** inline. Separately note that Rosetta and PyRosetta themselves are not OSS — they require a University of Washington licence (free for academic/non-commercial use), which the page states. Un-flag once upstream commits a LICENSE file.

• UniProt MCP Server (uniprot.md) — flagged 2026-08-01 on a self-contradictory upstream license, picked up from the verifier’s 2026-07-29 security: caution note during the link/pricing recheck. The committed LICENSE in Augmented-Nature/UniProt-MCP-Server is a restrictive non-commercial grant while package.json and the README claim MIT. The page’s Pricing row previously asserted “Free / OSS (MIT)” — corrected this run to “Free to use” plus an inline **Unverified —** on the redistribution terms, with a Notes paragraph telling readers the server is fine to run (thin read-only client over the public UniProt REST API) but should not be vendored or redistributed until upstream resolves the conflict. The CC BY 4.0 SciAgent-Skills packaging already on the page is the cleanly-licensed alternative. Last upstream push 2025-12-21. Un-flag once one license wins.

• GlyGen MCP Server (glygen.md) — flagged 2026-07-15 on an unspecified upstream license (same bar as ChemCP / PLSDB / GROMACS MCP). The glygener/glygen-mcp-server GitHub license field is null and no LICENSE file is present. The endpoint is a GlyGen-first-party hosted connector (installable today), so it is catalogued with Pricing marked Unverified — inline rather than asserting Free / OSS. Revisit once upstream declares an explicit license for the wrapper.

• anthropics/healthcare standalone plugins deprecated → consolidated healthcare@healthcare — flagged 2026-07-12, resolved 2026-07-19. As of the current .claude-plugin/marketplace.json (re-fetched 2026-07-19), the standalone plugin entries clinical-trial-protocol, prior-auth-review, fhir-developer, cms-coverage, npi-registry, pubmed, and icd10-codes are all labelled “Deprecated — install healthcare@healthcare instead” and the single consolidated healthcare plugin bundles the skills + connected MCPs. Confirmed hosted MCP endpoints from the plugin .mcp.json: cms_coverage, icd10_codes, npi_registry, clinical_trials at https://hcls.mcp.claude.com/<name>/mcp, and PubMed at https://pubmed.mcp.claude.com/mcp. Updated across two runs: prior-auth-review.md + icd-10-codes.md (2026-07-12); clinical-trial-protocol.md, fhir-developer.md, cms-coverage.md, npi-registry.md (2026-07-19 — install snippets now point at healthcare@healthcare; CMS Coverage + NPI gained the literal hosted HTTP endpoint + claude mcp add + Claude Desktop mcp-remote snippets). No change needed to pubmed.md: its install path uses the life-sciences marketplace (pubmed@life-sciences), which is not deprecated — the deprecation applies only to the separate pubmed plugin in the healthcare marketplace; the endpoint https://pubmed.mcp.claude.com/mcp matches. Un-flagged.

• GROMACS MCP Server (gromacs-mcp.md) — flagged 2026-07-11 on an unspecified upstream license (same bar as ChemCP / PLSDB / OpenClaw). The GitHub license field is null and no LICENSE file is present; the README’s “LGPL (GROMACS)” refers to GROMACS itself, not the MCP wrapper. Catalogued (Docker-based, followable install) with Pricing marked Unverified — inline rather than asserting Free / OSS. Revisit once upstream declares an explicit license for the wrapper; downgrade or remove the OSS-adjacent framing if the terms turn out restrictive.

• Open Targets Plugin (open-targets.md) — flagged 2026-06-20 per user report #42 (@goodb, verified 2026-06-15): the remote MCP endpoint https://mcp.platform.opentargets.org/mcp fails the MCP initialize handshake with JSON-RPC -32602 "Invalid request parameters" under protocolVersion 2025-06-18 and 2024-11-05, so no tools register and a restart does not help (server-side). Endpoint reachable (HTTP 200, text/event-stream) but non-compliant. Field report + two workarounds (direct GraphQL API; ToolUniverse OpenTargets_* tools) added to the page Notes. Un-flag once the upstream MCP server returns a compliant initialize response.
• OpenClaw-Medical-Skills (FreedomIntelligence/OpenClaw-Medical-Skills) — flagged 2026-06-11 on a license contradiction, blocking batch ingest. The README markets it as “a curated collection of open-source medical and bioinformatics skills,” but (a) there is no LICENSE file in the repo, and (b) 185 of 449 SKILL.md files carry a per-file proprietary header: “All Rights Reserved. This code is proprietary and confidential. Unauthorized copying of this file, via any medium is strictly prohibited.” (© 2026 MD BABU MIA, PhD, MSSM — “Universal Biomedical Skills” project). The remaining ~59 carry OSI license lines and the rest none. Generating “Free / OSS” catalog pages over a collection that is partly all-rights-reserved would be inaccurate and is not done. Do not ingest until upstream publishes a single coherent repo license (or per-skill licenses we can read programmatically). Note: most of OpenClaw’s life-science surface (it shares many skill slugs with K-Dense/SciAgent — adaptyv, anndata, alphafold, arboreto, …) is already catalogued from the two cleanly-licensed collections, so the unique-coverage loss from deferring is small.
• ChemCP (scottmreed/ChemCP) — catalogued 2026-06-13 but no LICENSE file is declared upstream; the page lists pricing as Free / OSS on the basis of the public npm package and marks the redistribution terms Unverified — inline. Revisit if upstream adds an explicit license, or downgrade the OSS claim if it turns out to be restricted.
• allenbrain-mcp — upstream license unset; flagged 2026-05-20 before any redistribution.
• biorxiv@life-sciences / clinical-trials@life-sciences plugins — flagged 2026-05-28; published in the anthropics/life-sciences marketplace per upstream commit e96556b but the backing mcp.deepsense.ai MCP host returns NXDOMAIN. No catalog entry created until the endpoint is restored.

Deferred — next-run priority

• ToolUniverse skills/ has grown well past the ~68 recorded — full re-listing needed (2026-08-16 Drug Repurposing and Discovery pass). The skills/ directory listing was pulled in two alphabetical halves this run and confirms substantial growth since the 2026-08-02 count. DRD-adjacent directories seen but not catalogued this run (soft cap): tooluniverse-kegg-disease-drug (KEGG disease↔drug workflow — check first whether it should instead augment kegg-database.md / kegg-pathway-analysis.md), tooluniverse-crispr-screen-analysis and tooluniverse-functional-genomics-screens (these two are the most interesting outstanding items — the catalog’s standing “no CRISPR / gRNA-design surface” gap is a design-tool gap, and a screen-analysis skill is a partial answer worth taking on the next MCB slot), tooluniverse-antibody-engineering (Immunology + DRD), tooluniverse-computational-biophysics, tooluniverse-cancer-classification, tooluniverse-clinical-risk-scoring, tooluniverse-diagnostic-test-evaluation, tooluniverse-clinical-guidelines, tooluniverse-clinical-data-integration, tooluniverse-acmg-variant-classification (likely overlaps the newly-catalogued acmg-classification.md bioSkill — compare before creating a second page), tooluniverse-epidemiological-analysis, tooluniverse-meta-analysis, tooluniverse-mendelian-randomization ✅ (done this run), tooluniverse-gene-liability ✅, tooluniverse-gpcr-structural-pharmacology ✅, tooluniverse-aging-senescence ✅, tooluniverse-enzyme-kinetics ✅. Also newly visible and worth a non-DRD slot: tooluniverse-hla-immunogenomics, tooluniverse-immune-repertoire-analysis, tooluniverse-infectious-disease, tooluniverse-microbiome-research, tooluniverse-metagenomics-analysis, tooluniverse-microbial-genome-characterization (Immunology); tooluniverse-epigenomics / tooluniverse-epigenomics-chromatin, tooluniverse-gene-regulatory-networks, tooluniverse-molecular-cloning, tooluniverse-model-organism-genetics, tooluniverse-multi-omics-integration, tooluniverse-comparative-genomics, tooluniverse-fastq-qc, tooluniverse-expression-data-retrieval (MCB); tooluniverse-electron-microscopy (note — this is the closest thing yet to the standing cryo-EM gap; check whether it wraps RELION/cryoSPARC or is purely an analysis reasoning layer); tooluniverse-lipidomics, tooluniverse-metabolomics / -metabolomics-analysis / -metabolomics-pathway; tooluniverse-literature-deep-research, tooluniverse-dataset-discovery, tooluniverse-data-wrangling, tooluniverse-biomedical-fact-lookup, tooluniverse-image-analysis (candidates for General-Purpose Utilities); and tooluniverse-ecology-biodiversity, tooluniverse-inorganic-physical-chemistry (adjacent-domain → utilities shelf). The devtu-* and create-tooluniverse-skill / setup-tooluniverse / tooluniverse-install-skills / tooluniverse-cs-setup / tooluniverse-custom-tool / tooluniverse-*-plugin directories are harness and authoring tooling, out of scope. Next ToolUniverse pass should capture the complete listing in one go and record the true count.
• GPCRdb, PDBePISA, SAbDab and OpenGenes have no discrete catalog pages (noted 2026-08-16 DRD pass). All four are now reachable through the newly-catalogued tooluniverse-gpcr-structural-pharmacology.md and tooluniverse-aging-senescence.md, but none has an entry of its own and none was checked this run for a standalone MCP/Skill wrapper. GPCRdb and SAbDab in particular have well-documented REST APIs that would map cleanly onto MCP tools; SAbDab is also referenced by the catalogued blatant-why.md (which bundles a sabdab MCP server inside its scaffold — that bundled server may be extractable, check 001TMF/blatant-why). Worth a targeted search on the next Structural or Immunology slot.
• MCPmed/STRINGmcp — augment, not a new page (surfaced 2026-08-15 Molecular and Cellular Biology pass, MCPmed org sweep). BSD-3-Clause, 3 stars, last push 2025-07-28, described upstream as “STRING DB mcp implementation” and named in the MCPmed Briefings in Bioinformatics 2026 paper as one of its three reference implementations alongside GEOmcp and UCSCCBmcp (both now catalogued). The catalog’s string-database-ppi.md does not mention MCPmed or STRINGmcp at all (grepped this run) — so per the one-entry-per-tool rule this is an alternative install path to fold into that existing page, not a new entry. Next MCB or Structural slot: fetch the repo README and pyproject.toml, confirm the PyPI package name actually resolves (the sibling UCSCCBmcp README advertised a PyPI package that turned out to 404, so check before publishing the command), enumerate the tools, and add an “Option — MCPmed STRING MCP server” block to string-database-ppi.md. Cheap, high-confidence work. Most relevant to Molecular and Cellular Biology.
• No CRISPR / gRNA-design MCP server or Skill exists — standing Molecular and Cellular Biology gap (confirmed 2026-08-15). A directed search plus a scan of the MCPmed hub found zero genome-editing MCP servers: the mature design tools are all pre-MCP research software with no Claude wrapper — crisprVerse/crisprDesign (R/Bioconductor; CRISPRko/a/i and base editors, on/off-target scoring, SNP and repeat annotation), USDA-ARS-GBRU/GuideMaker (Python CLI + web, HNSW-based, any Cas enzyme, genome-wide bacterial design in 1–2 min), GuideScan2, dylanbeeber/crispRdesignR, and the CRISPOR / CHOPCHOP / CRISPick webservers. The catalog’s only adjacent coverage is sgrna-design-guide.md (guidance, not a design engine) and plannotate-plasmid-annotation.md. Note the R angle: crisprDesign is Bioconductor, so the newly-catalogued mcptools.md is now a real route to it without anyone writing a server — worth saying so on sgrna-design-guide.md on a future pass. Re-check each MCB slot; davidliwei/awesome-CRISPR is the best single index to watch. Most relevant to Molecular and Cellular Biology.
• ClawBio/ClawBio — collection never ingested, now ships its own MCP server (re-surfaced 2026-08-15 MCB pass). Listed in AGENT.md as an authoritative community source but zero pages in the catalog trace to it. Upstream describes v0.5.0 (released 2026-04-04) as 95 skills (89 with runnable demo data) plus 8,182 Galaxy tools and 4,280 tests, with a no-install MCP server running locally over stdio and defaulting to demo data only. Deferred as batch-ingest work, not per-run work — at 95 skills it belongs in scripts/ingest_collection.py with a scripts/clawbio_category_map.yaml, following the K-Dense / SciAgent / NeuroClaw pattern, and would blow the ≤5-new-entries soft cap many times over. Before batching: confirm the repo license, check how much of the 95 duplicates already-catalogued tools (likely high overlap with the bioSkills and SciAgent sets — expect many augments rather than creates), and decide whether the bundled MCP server warrants one page of its own separate from the skills. Most relevant to Molecular and Cellular Biology.
• MCPmed/PLSDBmcp — located this run, still not catalogable (re-checked 2026-08-15 MCB pass; replaces the earlier 2026-08-15 Immunology note, now removed, which could not locate the repo at all). The repo does exist — 2 stars, last push 2025-06-26, described as “An MCP server linking the PLSDB database” — but the GitHub API reports license: None, i.e. no LICENSE file at all, and it is the only MCPmed repo without one (every sibling is BSD-3-Clause or MIT). Declined on the license bar, not on relevance: it would still be the catalog’s first plasmid-database surface and directly useful for bacterial AMR and mobile-element work. Revisit if upstream adds a license; the rest of the verification is cheap because the MCPmed packaging pattern is now well understood from geo-database.md, allenbrain.md and ucsc-cell-browser.md. Most relevant to Immunology and Microbiology.
• bioSkills alignment/ — 3 of 7 remaining (opened 2026-08-15, Integrative Structural and Computational Biology pass). Full 7-skill listing captured: alignment-io, alignment-trimming ✅, msa-parsing, msa-statistics ✅, multiple-alignment ✅, pairwise-alignment, structural-alignment ✅. The three left need an overlap audit before a page is created: pairwise-alignment (likely Bio.Align / parasail / EMBOSS — audit against biopython.md and blast.md), msa-parsing and alignment-io (format conversion; almost certainly collapse into biopython.md as an alt-install path rather than earning pages). Sibling category alignment-files/ (SAM/BAM/CRAM-side, judging by name) has no listing fetched yet and is read-alignment rather than MSA — fetch on a Molecular and Cellular Biology slot and route against samtools-bam-processing.md / pysam.md.
• bioSkills comparative-genomics/ — entirely uncatalogued, 13-skill listing captured 2026-08-15: ancestral-reconstruction, comparative-annotation-projection, gene-family-evolution, gene-tree-species-tree-reconciliation, genome-distance-and-species-delineation, hgt-detection, introgression-detection, ortholog-inference, pangenome-analysis, positive-selection, synteny-analysis, whole-genome-alignment, whole-genome-duplication. The catalog has no coverage of any of these (roary-pangenome.md is the only adjacent page and is bacterial-pangenome-specific). The two strongest leads are positive-selection (PAML/HyPhy dN/dS — the natural downstream consumer of the codon-alignment guidance now on multiple-alignment.md) and ortholog-inference (OrthoFinder/BUSCO-shaped); whole-genome-alignment (minimap2/MUMmer/Cactus) and synteny-analysis are the next tier. Most naturally tagged Integrative Structural and Computational Biology — lead with these on the next Structural slot.
• agentcures/refua-mcp — “Model Context Protocol server exposing typed Refua tools for folding, affinity, design, and optional clinical and preclinical workflows”, topics mcp, drug-discovery, clinical-trials, protein-folding, preclinical, antibody-design, affinity-prediction, refua. Surfaced 2026-08-15 Immunology and Microbiology pass via the GitHub antibody-design topic. Deferred on low signal: 0 stars, no license/package/endpoint confirmed from a primary source this run, and “Refua tools” is undefined outside the vendor’s own naming — it is unclear whether the backing service is public or a paid Agentcures account. Its claimed span (folding + affinity + antibody design + clinical/preclinical) would overlap the newly-catalogued blatant-why.md heavily. Next Immunology slot: read the README and package.json/pyproject.toml, confirm license and whether the tools call a hosted API needing a key; catalogue only if a copy-pasteable registration snippet and a license both exist. Most relevant to Immunology and Microbiology + Drug Repurposing and Discovery.
• BV-BRC (Bacterial and Viral Bioinformatics Resource Center) — no MCP surface exists (reconfirmed 2026-08-15 Immunology and Microbiology pass). 14M+ public pathogen genomes and 33 high-throughput analysis services, with open APIs and the p3-* CLI that would map nearly 1:1 onto MCP tool definitions; the 2026 NAR database paper is even titled “…with expanded functionality and AI integration”. No MCP server or Skill wraps it, and nothing catalogable resulted. This is the cleanest unfilled microbiology gap in the catalog — bacterial/viral pathogen genomics is currently reachable only through generic NCBI/assembly servers plus the local-tool skills (bakta-genome-annotation.md, prokka-genome-annotation.md, roary-pangenome.md, amr-detection.md, kraken-classification.md, metaphlan-profiling.md, mgnify.md). Revisit each Immunology slot; also re-check whether BV-BRC’s own “AI integration” ships as an MCP endpoint. IEDB is the matching immunology-side gap — its Query API and MHC-I/II + B-cell prediction tools still have no MCP/Skill wrapper (MCP Registry searches for epitope and antibody returned zero servers on 2026-08-15); it reaches Claude only indirectly through can-immune.md (NetMHCpan-EL via the IEDB API) and the tooluniverse-* skills.
• SYNTHIA® Agentic AI (Merck KGaA / Sigma-Aldrich, synthiaonline.com/product/synthia-agentic-ai) — the strongest dedicated retrosynthesis MCP surface located on the 2026-08-15 Chemistry pass, and the closest thing to a commercial answer to the standing retrosynthesis gap. Exposes the 21-year-old SYNTHIA expert-rule retrosynthesis engine over MCP alongside reaction annotations, chemical metadata, GHS/safety data and regulatory databases, with a catalog of 12M+ commercially available building blocks; deploys as a connector into the customer’s own LLM environment over a secure cloud connection, and the vendor states customer data is not used for training. Deferred 2026-08-15 — no public Claude install path. It is enterprise-only and quote-based (custom term lengths, unlimited users, no published seat/term price; G2 confirms pricing is not disclosed), and crucially no MCP endpoint URL, claude mcp add snippet, or claude.com/connectors listing is published — the go-to-market is a sales conversation, so there is nothing a reader could copy and run. That places it with the hosted-SaaS-without-a-Claude-installable-surface exclusions rather than with the subscription-gated-but-installable entries like cortellis.md (which has a real /plugin install path). Adjacent entry points noted for context: an individual free trial with business-email registration, SYNTHIA Lab for academic/non-profit teams (advertised 50% introductory discount, 10 full users, unlimited molecule runs), and an older pre-MCP SAS/retro REST API. Revisit if Merck publishes an endpoint URL or a connector listing — at that point it becomes catalogable as an Enterprise-priced Chemistry entry. Most relevant to Chemistry + Drug Repurposing and Discovery.
• ghutchis/chem-skill (Geoff Hutchison — Avogadro / Open Babel author, MIT via GitHub API, default branch main) — a small Claude Skill generating 2D structure images and interactive 3D conformer viewers from chemical names, SMILES or InChI, built on RDKit + pubchempy + py2opsin (offline IUPAC parsing), with three scripts (scripts/chem_2d.py, scripts/chem_3d.py, scripts/chem_vis.py). Requires network access to *.nih.gov for PubChem name resolution. Surfaced 2026-08-15 Chemistry pass. Deferred on the followable-verbatim bar plus low signal: the repo describes itself as “An experiment with Claude Skills”, the README’s install block clones a placeholder repo that does not exist (https://github.com/yourusername/chem-vis.git) and the project is inconsistently named chem-skill vs chem-vis; it documents pip install rdkit pubchempy py2opsin and direct python scripts/… invocation but never states how to install it into ~/.claude/skills/, so there is no verified skill-install path to publish. Also stale (last push 2025-12-28) at 3 stars. Coverage loss is small — 2D depiction is already served by cdxml-toolkit.md (publication-quality drawings, reaction schemes, curved-arrow mechanisms, ChemDraw CDXML export) and 3D viewing by the py3Dmol skill inside the newly-catalogued alkyl.md. Revisit if upstream fixes the clone URL and documents the skills-directory install. Most relevant to Chemistry.

• cyanheads/pubchem-mcp-server — a fourth PubChem MCP packaging (search by name, SMILES, InChIKey, formula, substructure and similarity, plus physicochemical properties and classification; STDIO or Streamable HTTP, and a public hosted instance requiring no installation). Surfaced 2026-08-15 Chemistry pass. Not catalogued as a new entry — candidate augment to the existing pubchem.md rather than a duplicate, per the one-entry-per-tool rule. The same author’s protein-mcp-server, openfda-mcp-server and clinicaltrials-gov-mcp are already catalogued and have proven reliable (Apache-2.0, npm-published, public HTTP instances), so this is likely a clean add. Next Chemistry slot: verify the license, npm package name, tool list and hosted endpoint URL, then fold the install path into pubchem.md — do not create a new page. Note the catalog also already carries PhelanShao/pubchem-mcp-server and JackKuo666/PubChem-MCP-Server variants in the wild; check which, if any, pubchem.md currently documents before editing. Most relevant to Chemistry.

• bioSkills clinical-biostatistics/ + clinical-databases/ — two newly-listed Translational Medicine categories (listings confirmed 2026-08-02). clinical-biostatistics/ — DONE as of 2026-08-09, nothing left worth cataloguing. All 12: adaptive-designs ✅, bayesian-trials ✅, categorical-tests ✗, cdisc-data-handling ✅, effect-measures ✅, logistic-regression ✗, missing-data-sensitivity ✅, multiplicity-graphical ✅, power-and-sample-size ✅, subgroup-analysis ✅, survival-analysis ✗, trial-reporting ✅ (three catalogued 2026-08-02, six on 2026-08-09). The three ✗ are deliberate declines: survival-analysis overlaps scikit-survival.md — fold the bioSkills install path into that page rather than duplicating; logistic-regression and categorical-tests are generic statistics that belong on the General-Purpose Utilities shelf if anywhere, and are already covered in substance by statsmodels.md. Do not re-surface this category; diff-only from here. See the CDISC Library MCP entry below for the metadata server that complements the now-catalogued cdisc-data-handling.md. clinical-databases/ (12 skills) — four of the five genuinely-distinct ones catalogued 2026-08-16: acmg-classification ✅, clinvar-lookup ✗, dbsnp-queries ✗, gnomad-frequencies ✗, hla-typing ✗, msi-detection ✅, myvariant-queries ✗, pharmacogenomics ✗, polygenic-risk (only one left — do this next), somatic-signatures ✅, tumor-mutational-burden ✅, variant-prioritization ✗. The seven ✗ wrap data sources already catalogued (ClinVar, dbSNP, gnomAD, ClinPGx) or generic prioritization workflow — fold the bioSkills install path into the existing pages rather than creating duplicates. polygenic-risk is verified and ready to write (front-matter read 2026-08-16: bio-clinical-databases-polygenic-risk, tool_type: mixed, primary_tool: PGS Catalog Calculator): LDpred2-auto (bigsnpr), PRS-CS/PRS-CSx, SBayesRC (GCTB), MegaPRS (LDAK), multi-ancestry PROSPER / MUSSEL / BridgePRS / JointPRS / PRSmix, pgsc_calc (nf-core Nextflow) as the production pipeline, PRSice-2 and lassosum2 as legacy; ancestry-conditional Z-score calibration against test-cohort PCs (not discovery PCs), LD-mismatch s < 0.05, sample-overlap detection via EraSOR or bivariate LDSC (

  intercept

  > 0.05 at n ≥ 1000), Martin 2017 / Ding 2023 cross-ancestry portability, and the 22-item PRS-RS reporting checklist. Route it against plink2-gwas-analysis.md and tooluniverse-gwas-drug-discovery.md for cross-links; deferred this run only by the ≤5 soft cap. Also worth noting: MyVariant.info has no catalog page at all (mygene.md exists but no MyVariant equivalent), so myvariant-queries may in fact be an uncovered surface rather than a duplicate — check before dismissing it.

• CDISC Library MCP Server (Cicatriiz/cdisc-library-mcp-server, MIT via GitHub API, Node >= 18) — the only located MCP surface over the CDISC Library: generic list_operations / call_operation, discovery search.suggest / search.query, Biomedical Concepts bc.list / bc.get / bc.categories (+ package variants), and SDTM Dataset Specializations sdtm.list / sdtm.get / sdtm.domains / sdtm.byBiomedicalConcept (+ package variants), plus cache management. Requires a CDISC_API_KEY subscription key from api.developer.library.cdisc.org (free CDISC Library account at library.cdisc.org/browser), passed via header (default) or query per CDISC_AUTH_LOCATION. Surfaced 2026-08-02 Translational Medicine pass. Deferred on the followable-verbatim bar: the README documents only export CDISC_API_KEY=… && npm start from a clone plus dxt pack for Claude Desktop — there is no published npm package, no stated server entry-point path for a literal claude mcp add --transport stdio … -- node <path> snippet, and the .dxt extension format has since been superseded by MCPB, so the Desktop path may no longer install cleanly. Repo is also stale (last push 2025-08-12, 4 stars). Revisit once the entry point / an npm package / an MCPB bundle is confirmed. Complementary to the bioSkills cdisc-data-handling skill above. Most relevant to Translational Medicine.
• OMCP (FastOMOP) (fastomop/omcp, KCL/UCL) — a governed natural-language-to-SQL MCP layer over an OMOP CDM instance, with an authenticated API tier, SQLGlot + Ibis transpilation across backends, and an output validator that anonymises or errors on identifiable data; the OHDSI 2025 evaluation reported Claude Sonnet at a 93.3% success rate (4.73 attempts/query). Surfaced 2026-08-02 Translational Medicine pass alongside the catalogued omophub-mcp.md / pyomop.md. Deferred: the license and a copy-pasteable Claude registration snippet were not confirmed from a primary source this run, and the “servers” (plural) framing suggests a multi-component deployment rather than one installable server. Its distinguishing value over pyomop.md is the governance/de-identification layer, which is exactly what a hospital deployment would want. Revisit next Translational Medicine slot. Most relevant to Translational Medicine.
• ChemGraph HPC / Parsl sibling MCP modules (argonne-lcf/ChemGraph, Apache-2.0) — the base server is now catalogued as chemgraph.md (2026-08-08). The same package ships six further MCP modules that are launched identically by swapping the module name: mace_mcp_parsl, graspa_mcp_parsl (gRASPA gas adsorption), xanes_mcp_parsl (XANES spectra), ase_mcp_hpc, data_analysis_mcp, hpc_misc_mcp (plus transfer_tools / job_tools helpers). Deferred: none is documented upstream — no README in src/chemgraph/mcp/, no enumerated tool list, and the Parsl/HPC variants imply a scheduler configuration that is not published. They are named on chemgraph.md as unverified. Revisit if upstream documents them; graspa_mcp_parsl (adsorption) and xanes_mcp_parsl (spectroscopy) would each be a genuinely new catalog surface — the latter would partly answer the deferred spectroscopy gap below. Most relevant to Chemistry.
• DELFIN (ComPlat/DELFIN, LGPL-3.0) — quantum-chemistry agent layer with MCP support wrapping 11 ab initio/DFT programs (ORCA, Gaussian, TURBOMOLE, NWChem, Q-Chem, GAMESS, Molpro, Dalton, Psi4, CFOUR, MRCC) plus xTB/CREST/MOPAC. Surfaced 2026-08-01 Chemistry pass. Deferred: the MCP surface is described as part of a broader agent harness rather than a discrete installable MCP server, and the LGPL-3.0 grant explicitly conveys no rights to the wrapped engines (ORCA, xTB, CREST each carry their own terms), so the pricing/licensing row would need careful per-engine framing. Revisit once the MCP server is documented as a standalone install with a copy-pasteable registration snippet. Most relevant to Chemistry.
• GPAW Computation MCP (pathintegral-institute/mcp.science, MIT) — density-functional-theory server in the same collection as the newly-catalogued materials-project-mcp.md. Deferred 2026-08-01: fails the followable-verbatim bar — the README enumerates no tools, requires installing a server_package folder on a remote machine, references an unexplained “MPI API key”, and depends on the Materials Project MCP’s output directory via a hand-edited settings.toml. Revisit once the remote-server setup is documented end-to-end. Would be Chemistry (or General-Purpose Utilities if framed as materials DFT).
• s20ss/mcp_rdkit — third RDKit MCP packaging (molecular visualization + descriptor calculation, stdio via python -m mcp_rdkit). Surfaced 2026-08-01 Chemistry pass. Deferred as redundant: the catalog already carries three RDKit entries — rdkit-mcp.md (TandemAI, full RDKit surface as MCP tools), rdkit-skill.md (K-Dense Skill), and rdkit-agent.md (WASM CLI/MCP) — and this variant covers a strict subset. Fold in as an alternative install path on rdkit-mcp.md only if its license and tool list are later confirmed and it offers something the TandemAI server does not.
• NMR / tandem-MS spectra-prediction MCP — no Claude-installable wrapper located on the 2026-08-01 Chemistry pass. The strong models ship as plain Python research artifacts (out of scope as bare installs): coleygroup/ms-pred (ICEBERG 2.1 neural tandem-MS prediction, SCARF formula-level spectra) and rxn4chemistry/nmr-to-structure (transformer structure elucidation from NMR). The only Claude-reachable spectroscopy surface today is the commercial hosted covasyn.md (mass spec + NMR among its 130+ functions) and the newly-catalogued xtb-mcp-server.md (generate_spectroscopy_input, input prep only). Revisit if an OSS spectra-prediction MCP ships. Most relevant to Chemistry.
• BrainGlobe atlas-registration stack (brainglobe-atlasapi + DeepSlice) — user request #74 (@goodb, 2026-07-27, via recipes-block chain) needs these two for a QUINT-replacement implant-localization recipe. Both verified real and cleanly licensed this run (bank these — no re-verification needed): brainglobe-atlasapi is BSD-3-Clause (JOSS 2020, pip install brainglobe-atlasapi, Python 3.9+, ships allen_mouse_25um with annotation/lookup_df/structures), DeepSlice is GPL-3.0-only (PyPI license_expression for 1.2.8, re-verified 2026-07-27 — an earlier pass in this same request recorded MIT, which was wrong; copyleft, so redistribution of a derived pipeline carries obligations; Carey et al., Nat Commun 2023; pip install DeepSlice, coronal-only, QuickNII-compatible output). Declined to catalogue 2026-07-27 — out of scope as bare PyPI libraries: neither ships as a Claude-installable Skill / MCP / plugin / connector (checked PyPI, both repos, and the NeuroClaw / SciAgent / K-Dense / bioSkills collections — no SKILL.md or MCP wrapper exists for either). Same bar as raw RDKit/Scanpy. Revisit the moment a community SKILL.md or MCP wrapper for either surfaces — the license/identity verification is done, so cataloguing is then a fast create + the recipes rung-3 recipe unblocks. Most relevant to Neuroscience.
• openpharma-org/fda-mcp — RESOLVED and catalogued 2026-08-09 as fda-mcp.md (Drug Repurposing and Discovery + Translational Medicine). The 2026-07-26 blocker (unspecified license) cleared this run: the GitHub license API now returns MIT with a LICENSE file in the repo root. Correction to the original deferral note: there is no npm package — both fda-mcp and the package.json name @openpharma-org/fda-mcp-server 404 on the npm registry (checked 2026-08-09) — so the page installs from source (git clone → npm install → npm run build → register node /path/to/fda-mcp/build/index.js), and it warns that the upstream README’s /path/to/fda-mcp-server/… path does not match the repo’s own directory name fda-mcp. package.json verified: v1.0.0, main/bin → build/index.js, engines.node >=18.0.0, build: "tsc". Availability recorded as Beta (source-only install, repo pushed 2026-03-10, 3 stars). Unique surface confirmed as the reason to carry it: Orange Book patent/exclusivity + Purple Book biosimilar interchangeability, neither covered by openfda.md / openfda-mcp-server.md. Sibling openpharma-org/medicare-mcp was NOT re-checked this run — check its license next Translational Medicine slot; the org has now declared one license, so the sibling may have too.

• openpharma-org/medicare-mcp — RESOLVED and catalogued 2026-08-16 as medicare-mcp.md (Translational Medicine). The 2026-07-26 blocker (unspecified license) cleared: the GitHub license API now returns MIT with a LICENSE file in the repo root. Correction to the original deferral note: there is no npm package — both medicare-mcp and the package.json name @openpharma-org/medicare-mcp-server 404 on the npm registry (checked 2026-08-16) — so the page installs from source (git clone → npm install → npm run build → register node /path/to/medicare-mcp/build/index.js) and warns that the upstream README’s /path/to/medicare-mcp-server/… path does not match the repo’s own directory name medicare-mcp. package.json verified: v0.3.0, bin → build/index.js, engines.node >=16.0.0. Availability recorded as Beta (source-only install). The 17 methods were confirmed from the README (one more than the 16+ recorded in the original note). Original note follows. CMS Medicare MCP via a unified medicare_info tool with 16+ methods: provider/prescriber/hospital search, hospital star ratings, readmission/mortality/infection rates, VBP/HCAHPS scores, ASP drug pricing + trends (2013–2023, Socrata API), keyless, Node/TS. Surfaced 2026-07-26 Translational Medicine pass. Deferred on an unspecified upstream license (same org / same gap as fda-mcp). Complements the newly-catalogued cms-datagov-mcp.md (dataset catalog) with pre-shaped Medicare quality/pricing tools. Revisit once a license is declared. Most relevant to Translational Medicine.

• stMCP — spatial-transcriptomics MCP (bioRxiv 2026.03.11.710153, Case Western Reserve BME) — a Claude-Desktop-targeting spatial-transcriptomics MCP that ships a single self-contained setup script auto-registering the server with Claude Desktop (Conda-forge-bundled portable runtime, local execution to avoid uploading large datasets). Surfaced on the 2026-07-18 Molecular and Cellular Biology directed pass (spatial transcriptomics MCP server query). Deferred: the exact GitHub repository URL and software license could not be verified from a primary source (the preprint’s Code Availability section was not retrievable via search; no repo/license surfaced). Revisit once the repo URL + license + a copy-pasteable install snippet are confirmed. Complements the catalogued chatspatial.md (60+ spatial methods). Most relevant to Molecular and Cellular Biology.
• HughYau/neuroforge-skills — collection now exhausted of new-to-catalog skills — a small MIT-licensed neuroscience skill set of 5 skills (flat skills/<slug>/, install via git clone + cp -r neuroforge-skills/skills/<slug> ~/.claude/skills/ plus the underlying pip install). Batch-covered in diff-only terms 2026-07-12: Brian2 (brian2.md) and pyNIBS (pynibs.md) catalogued this run; the other three wrap tools already catalogued — MNE-Python (NeuroClaw mne-eeg-tool.md), Nilearn (nilearn-tool.md), SpikeInterface (spikeinterface-electrophysiology.md). Fold the NeuroForge install path into those three only if a future run audits and confirms it. No genuinely new NeuroForge entries remain.
• HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills — audited 2026-08-02, now in incremental mode (surfaced 2026-07-12 via InternScience/Awesome-Scientific-Skills). Audit result: repo-level MIT (SPDX via GitHub API — no conflicting per-skill license headers found in the five skills read); default branch master (raw main URLs 404 — use master); 34 stars; last push 2026-07-13. Install is a genuine Claude Code plugin marketplace: /plugin marketplace add HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills then /plugin install awesome-cognitive-and-neuroscience-skills@awesome-cognitive-and-neuroscience-skills (marketplace name == plugin name, source: "./"), plus single-skill cp -r .../skills/<slug> ~/.claude/skills/. Skills are description-activated — no slash commands, so followable-verbatim rule 1 does not apply. Quality caveat that must appear on every page from this collection: all skills carry review_status: ai-generated and the README says the content “has not been individually verified by human domain experts” — each catalogued page leads its Notes with a bolded AI-generated caveat plus the skill’s cited papers. Dependency gotcha: most non-meta skills declare dependencies: required: [research-literacy], so the single-skill copy path must also copy skills/research-literacy. 50 directories, ~10 meta (contribute-skill, contribute-skills-via-pr, verify-skill, paper-to-skill, repo-to-skill, share-case, share-usage, research-literacy, markdown-report-writing — out of scope as curation/harness skills). Catalogued 2026-08-02: deeplabcut, calcium-imaging-analysis-guide, drift-diffusion-model, optogenetics-protocol-designer, signal-detection-analysis. Catalogued 2026-08-09 (50-directory listing re-confirmed unchanged): neural-population-analysis-guide, parameter-recovery-checker, lesion-symptom-mapping-guide, netneurotools-guide, pycortex-guide. Dependency finding from the 2026-08-09 batch: the dependencies: required: [research-literacy] block is not universal — it appears on the skills authored by “Claude Code Agent” (which also carry a papers: list and a domain: field) but not on the “Claude (AI-assisted)” ones (netneurotools-guide, pycortex-guide), which additionally cite no papers at all. Read the front-matter per skill rather than assuming; pages for the paperless subset should point readers at the wrapped library’s own documentation instead of a citation list. Strong remaining candidates with no catalog coverage: act-r-model-builder, bayesian-cognitive-model-builder, evidence-accumulation-selector (route against the now-catalogued drift-diffusion-model.md first — it may be a selector wrapper rather than a distinct surface), neuropsych-battery-selector, cogsci-power-analysis / neuroimaging-power-guide / neuroimaging-sample-size-calculator (three power/sample-size skills that likely collapse into one page — audit before creating three), cogsci-statistics / cogsci-visualization (probably General-Purpose Utilities if catalogued at all), divergent-thinking-scoring, creativity-self-efficacy-mediation, fmri-task-design-guide, signal-adjacent paradigm designers (eeg-paradigm-designer, cognitive-paradigm-design, visual-search-array-generator, infant-looking-time-designer, self-paced-reading-designer, sentence-stimulus-norming, tom-task-selector, alternative-uses-task-designer, reading-time-analysis). Overlap-audit-first (wrap tools already catalogued — fold in as alternative install paths rather than duplicating): fmriprep → fmriprep-tool.md, mne-python-guide → mne-eeg-tool.md, spikeinterface-skill → spikeinterface-electrophysiology.md, fmri-preprocessing-pipeline-guide / fmri-glm-analysis-guide → fmri-skill.md / glm.md, eeg-preprocessing-pipeline-guide / erp-analysis → eeg-skill.md, brain-connectivity-modeler → combraintf.md / bnt.md, neural-decoding-analysis → svm.md / nilearn-tool.md, spiking-network-model-builder → brian2.md. Surface 1–5 per future Neuroscience pass under the soft cap.
• GPTomics/bioSkills — batch-ingest candidate collection (surfaced 2026-07-11, Immunology and Microbiology directed pass) — a large MIT-licensed Claude/OpenCode skill library: 561 skills across 63 root-level category directories (no skills/ prefix — categories like immunoinformatics/, tcr-bcr-analysis/, metagenomics/, microbiome/, flow-cytometry/, imaging-mass-cytometry/, single-cell/, variant-calling/, … sit at repo root, each holding several skills). Install: git clone + ./install-claude.sh --categories "<cat>" (target ~/.claude/skills/) or single-skill cp -r <cat>/<skill> ~/.claude/skills/. Skill front-matter names are bio-<category>-<skill>. Catalogued so far: epitope-prediction, mhc-binding-prediction, immcantation-analysis, kraken-classification, amr-detection (2026-07-11); neoantigen-prediction, mhc-class-ii-prediction, scirpy-analysis, mixcr-analysis, tcr-epitope-binding (2026-07-18); repertoire-visualization, specificity-annotation, vdjtools-analysis, amplicon-processing, taxonomy-assignment (2026-07-25); metaphlan-profiling, functional-profiling, strain-tracking, immunogenicity-scoring, differential-abundance (2026-08-01); structure-validation, binding-site-detection, interface-analysis, structure-preparation (2026-08-01, Structural slot); doublet-detection, cnv-inference, perturb-seq, scatac-analysis, lineage-tracing (2026-08-01, MCB slot). Remaining single-cell/ candidates (full 17-skill listing confirmed 2026-08-01: batch-integration, cell-annotation, cell-communication, clustering, cnv-inference, data-io, differential-abundance, doublet-detection, hashing-demultiplexing, lineage-tracing, markers-annotation, metabolite-communication, multimodal-integration, perturb-seq, preprocessing, scatac-analysis, trajectory-inference) — the 12 not yet catalogued mostly overlap existing pages: preprocessing / data-io overlap single-cell-rna-qc.md / scanpy.md / anndata.md, cell-annotation + markers-annotation overlap celltypist-cell-annotation.md / popv-cell-annotation.md / single-cell-annotation-guide.md, cell-communication overlaps cellchat-cell-communication.md / liana-mcp.md, batch-integration overlaps harmony-batch-correction.md / scvi-tools.md, multimodal-integration overlaps muon-multiomics-singlecell.md / mofaplus-multi-omics.md, trajectory-inference overlaps scvelo.md / cellrank-mcp.md. Both previously-flagged genuinely-distinct single-cell/ skills were catalogued 2026-08-08: hashing-demultiplexing → hashing-demultiplexing.md and metabolite-communication → metabolite-communication.md. Still remaining there: clustering and single-cell differential-abundance (Milo/scCODA — note the slug collision with the already-catalogued metagenomics differential-abundance.md, so a distinct slug like sc-differential-abundance would be needed) warrant an overlap audit first. Remaining structural-biology/ candidates (full 10-skill listing confirmed 2026-08-01: alphafold-predictions, binding-site-detection, geometric-analysis, interface-analysis, modern-structure-prediction, structure-io, structure-modification, structure-navigation, structure-preparation, structure-validation) — geometric-analysis was catalogued 2026-08-08 (geometric-analysis.md) after an overlap audit found it genuinely distinct from the trajectory pages — it is the static-structure measurement layer (Bio.PDB distances/dihedrals/superposition/SASA), where mdtraj-trajectory-analysis.md / mdanalysis-trajectory.md are trajectory-side. The five still uncatalogued mostly overlap existing pages: alphafold-predictions and modern-structure-prediction overlap alphafold.md / boltz.md / chai-1.md / openfold3.md, and structure-io / structure-navigation / structure-modification overlap biopython.md; audit for genuinely-distinct content before creating pages, or fold in as alt-install paths. Sibling rna-structure/ category mined 2026-08-08 and now exhausted: full 4-skill listing confirmed (covariation-analysis, ncrna-search, secondary-structure-prediction, structure-probing) and three catalogued — covariation-analysis.md, ncrna-search.md, structure-probing.md. The fourth, secondary-structure-prediction, was deliberately not catalogued because it overlaps viennarna-structure-prediction.md (SciAgent packaging of the same library) — fold the bioSkills install path into that page if a future run audits it, rather than creating a second RNA-folding entry. Remaining strong immunology/microbiology candidates: metagenomics/ → abundance-estimation, contamination-controls, metagenome-visualization (full 8-skill listing confirmed 2026-08-01: abundance-estimation, amr-detection, contamination-controls, functional-profiling, kraken-classification, metagenome-visualization, metaphlan-profiling, strain-tracking); microbiome/ → diversity-analysis, qiime2-workflow, functional-prediction (full 6-skill listing confirmed 2026-08-08: amplicon-processing ✅, differential-abundance ✅, diversity-analysis, functional-prediction, qiime2-workflow, taxonomy-assignment ✅ — so those three are the only uncatalogued ones left in the category and are the cleanest next-run Immunology/Microbiology adds). flow-cytometry/ — mined 2026-08-08, now near-exhausted: full 8-skill listing confirmed (bead-normalization, clustering-phenotyping, compensation-transformation, cytometry-qc, differential-analysis, doublet-detection, fcs-handling, gating-analysis) and five catalogued this run — compensation-transformation.md, cytometry-qc.md, gating-analysis.md, clustering-phenotyping.md, cytometry-differential-analysis.md. Three deliberately not catalogued: fcs-handling overlaps the catalogued flowio.md (K-Dense FCS parser) — fold the bioSkills path into that page rather than duplicating; doublet-detection has a slug collision with the already-catalogued single-cell doublet-detection.md, so it would need a distinct slug like cytometry-doublet-detection and an overlap audit (the flow-cytometry version is FSC-A-vs-FSC-H/Event_length-based, genuinely different from scDblFinder/Scrublet, so it is a legitimate future add); bead-normalization (CyTOF EQ-bead signal-drift normalization) is genuinely distinct with no catalog coverage and is the strongest remaining single candidate in the category. hi-c-analysis/ — opened 2026-08-08, 6 of 9 remaining: full 9-skill listing confirmed (compartment-analysis ✅, contact-pairs, hic-data-io, hic-differential, hic-visualization, hichip-plac-loops, loop-calling ✅, matrix-operations, tad-detection ✅) and the three feature-calling layers catalogued. The six left are all genuinely new catalog surfaces (the catalog had zero 3D-genome coverage before this run): hichip-plac-loops (FitHiChIP/MAPS protein-anchored loops — explicitly scoped out of loop-calling.md, so it is the cleanest next add), hic-differential (cross-condition contact comparison), contact-pairs + hic-data-io + matrix-operations (the pairtools/cooler upstream layer — audit whether they collapse into one page or belong as a single “cooler/pairtools” entry rather than three), hic-visualization (route against deeptools.md / plotly-interactive-plots.md first). epitranscriptomics/ listing captured 2026-08-08 and entirely uncatalogued (5 skills: m6a-differential, m6a-peak-calling, m6anet-analysis, merip-preprocessing, modification-visualization) — RNA modification has no catalog coverage at all, so this is the single strongest untouched MCB category; m6a-peak-calling (MeRIP/exomePeak2-shaped) and m6anet-analysis (direct nanopore detection) are the two to lead with. The 65-directory root listing was re-fetched 2026-08-08 — other MCB-relevant categories still entirely unmined: alternative-splicing, methylation-analysis, ribo-seq, long-read-sequencing, spatial-transcriptomics, hi-c-adjacent gene-regulatory-networks, crispr-screens, liquid-biopsy, phasing-imputation, temporal-genomics, causal-genomics, epidemiological-genomics. imaging-mass-cytometry/ still entirely unmined — no listing fetched yet; expect Steinbock/IMC-Segmentation-Pipeline/imcRtools-shaped skills with zero catalog coverage, so fetch that listing first on the next Immunology and Microbiology slot. Recommended: a one-time batch ingest (like K-Dense/SciAgent/NeuroClaw) via scripts/ingest_collection.py + a bioskills_category_map.yaml, dedup-augmenting the ~dozens of slugs that overlap tools already catalogued (many bioinformatics skills mirror K-Dense/SciAgent coverage — dedup before creating). Until then, surface 1–5 genuinely-new skills per relevant directed pass under the soft cap. Note: many non-life-science-relevant category dirs (machine-learning, data-visualization, reporting, workflow-management) are General-Purpose Utilities or out of scope.
• Geneformer / scGPT single-cell foundation-model wrapper — no Claude-installable MCP server or Skill wrapping the single-cell RNA-seq foundation models Geneformer (ctheodoris/Geneformer, HF / NVIDIA BioNeMo) or scGPT (bowang-lab/scGPT, PyPI) located on the 2026-07-11 Molecular and Cellular Biology directed pass (Geneformer / scGPT MCP server Claude skill query). Both remain raw model weights + Python packages (research artifacts, out of scope as bare installs). Revisit if a discrete Claude Skill / MCP wrapper for zero-shot embedding / cell-type annotation surfaces. Most relevant to Molecular and Cellular Biology.
• GoekeLab/awesome-genomic-skills — discovery meta-list — a curated list of genomics/bioinformatics agent skills, MCPs, and benchmarks for Claude Code / Copilot / Codex / Cursor / Gemini CLI (surfaced 2026-07-04 on the Molecular and Cellular Biology directed pass). Most standalone entries are already catalogued (chatspatial, biomcp, gget, biocontextai/knowledgebase-mcp, clair-variant-caller); Seqera MCP was the genuinely-new find and is now catalogued (seqera.md). Remaining candidates to audit on future runs: the openai/plugins life-science-research collection (50 skills — confirm whether any are Claude-installable via the Agent Skills spec vs. OpenAI-only before ingesting) and a Clair-skills dedicated agent skill (HKU-...) that may be distinct from the catalogued clair-variant-caller packaging. Not a collection to batch-ingest itself — use it as a source index.
• PLSDB MCP (MCPmed/PLSDBmcp, PyPI plsdb-mcp) — MCP server over PLSDB (plasmid database): 8 tools (get_plasmid_summary, start_fasta_download / get_fasta_download, start_sequence_search / get_sequence_search_results, filter_plasmids_by_nuccore / _by_biosample / _by_taxonomy); no API key. Surfaced 2026-06-27 on the Molecular and Cellular Biology directed pass (MCPmed family). Deferred on an unspecified upstream license — the GitHub license field is null and no LICENSE file is shown, so a “Free / OSS” claim cannot be made (same bar applied to ChemCP / OpenClaw). Sibling geo-mcp (BSD-3-Clause) from the same org was catalogued (folded into geo-database.md). Revisit once PLSDBmcp declares an explicit license. Most relevant to Immunology and Microbiology / Molecular and Cellular Biology.

• MCPmed STRING DB / UCSC Cell Browser MCPs — two further MCPmed servers (Briefings in Bioinformatics 2026, bbag076). STRING-DB MCP overlaps the catalogued string-database-ppi entry (fold in as alt-install once verified); the UCSC Cell Browser MCP is reported “limited in function” upstream (no full API). Revisit per data source.

• google-deepmind/science-skills — collection now exhausted of new-to-catalog skills — the collection (Apache-2.0 code / CC-BY-4.0 docs, npx skills add google-deepmind/science-skills/ for Gemini/Antigravity; git clone + cp -r skills/<slug> ~/.claude/skills/ for Claude, plus the shared scienceskillscommon helper) batch-ingested in diff-only mode 2026-06-20. Catalogued: AlphaGenome / GTEx / UniBind / foldseek_structural_search (2026-06-27, foldseek-structural-search.md) / pymol (2026-06-27, pymol.md) / embl_ebi_ols (2026-06-27, embl-ebi-ols.md, the last genuinely new skill). The remaining skills are wrappers over data sources already catalogued under their own entries — protein_sequence_msa / protein_sequence_similarity_search (blast/uniprot), ncbi_sequence_fetch (NCBI / gene-database), ucsc_conservation_and_tfbs (ucsc-genome-browser), encode_ccres_database (encode-database) — fold in as alt-install paths only if a future run audits and confirms the existing pages should carry the science-skills install path. No genuinely new science-skills entries remain to surface.

• ToolUniverse sibling agent skills — mims-harvard/ToolUniverse/skills/ ships ~68 discrete Claude agent skills sharing the npx skills add mims-harvard/ToolUniverse install path (Apache-2.0). Catalogued so far: tooluniverse-drug-repurposing (2026-06-14, drug-repurposing.md), tooluniverse-drug-research / tooluniverse-drug-target-validation / tooluniverse-drug-synergy (2026-06-21), tooluniverse-drug-drug-interaction / tooluniverse-precision-oncology (2026-06-28), tooluniverse-pharmacovigilance / tooluniverse-rare-disease-diagnosis (2026-07-05, Translational slot), tooluniverse-target-research / tooluniverse-gwas-drug-discovery / tooluniverse-binder-discovery (2026-07-05, Drug Repurposing slot), tooluniverse-cancer-variant-interpretation / tooluniverse-adverse-event-detection / tooluniverse-immunotherapy-response-prediction / tooluniverse-adverse-outcome-pathway (2026-07-12, Drug Repurposing slot); tooluniverse-small-molecule-discovery / tooluniverse-chemical-safety / tooluniverse-cancer-genomics-tcga / tooluniverse-rare-disease-genomics (2026-07-19, Drug Repurposing slot); tooluniverse-admet-prediction / tooluniverse-dose-response / tooluniverse-cell-line-profiling / tooluniverse-chemical-sourcing (2026-08-02, Drug Repurposing slot); tooluniverse-drug-mechanism-research / tooluniverse-drug-regulatory / tooluniverse-clinical-trial-matching / tooluniverse-clinical-trial-design (2026-08-09, Drug Repurposing slot). Full skills/ listing re-confirmed 2026-08-09 — ~15 of the directories are harness/dev skills (create-tooluniverse-skill, setup-tooluniverse, tooluniverse-install-skills, tooluniverse-cs-setup, tooluniverse-custom-tool, and the devtu-* family) and are out of scope as curation tooling, as are the three IDE-packaging directories surfaced this run (tooluniverse-antigravity-plugin, tooluniverse-claude-code-plugin, tooluniverse-codex-plugin — distribution wrappers, not workflows; the Claude Code one may be worth a look as an alternative install path on tooluniverse.md). Strong remaining drug-discovery / clinical candidates: tooluniverse-cancer-classification, tooluniverse-clinical-guidelines, tooluniverse-clinical-data-integration, tooluniverse-comparative-genomics, tooluniverse-gwas-finemapping, tooluniverse-structural-variant-analysis, tooluniverse-clinical-risk-scoring, tooluniverse-diagnostic-test-evaluation, tooluniverse-aging-senescence, tooluniverse-acmg-variant-classification (audit against clinvar-database.md / clingen.md first), tooluniverse-disease-research, tooluniverse-chemical-compound-retrieval (likely overlaps pubchem.md / tooluniverse-chemical-sourcing.md — audit before creating). Other-area candidates: tooluniverse-antibody-engineering (Immunology), tooluniverse-crispr-screen-analysis (MCB), tooluniverse-computational-biophysics (Structural). Overlap-audit-first / likely out of scope: tooluniverse-data-wrangling, tooluniverse-dataset-discovery, tooluniverse-data-integration-analysis, tooluniverse-biomedical-fact-lookup. Each requires the ToolUniverse MCP server. Surface 1–5 per future Drug Repurposing / Translational pass; do not bulk-create (soft cap).
• Cicatriiz/healthcare-mcp-public (MIT via GitHub API, 125 stars) — a single MCP server bundling FDA drug info, PubMed, medRxiv, NCBI Bookshelf, clinical-trials search, ICD-10 lookup, DICOM metadata, and a medical calculator. Surfaced 2026-08-02 Drug Repurposing and Discovery pass. Deferred on overlap + staleness: the FDA / PubMed / clinical-trials / ICD-10 surfaces are each already catalogued discretely (openfda.md, openfda-mcp-server.md, pubmed.md, clinicaltrials-gov-mcp.md, icd-10-codes.md), leaving only medRxiv, NCBI Bookshelf, DICOM metadata, and the calculator as unique, and the last upstream push was 2025-08-16 (~1 year stale). Its real appeal is being one install instead of five — catalogue it as a single-install alternative if users ask, alongside the similarly-deferred easysolutions906 healthcare bundle. Most relevant to Translational Medicine.
• cyanheads/pubchem-mcp-server — PubChem MCP with stdio + streamable-HTTP transports, adding drug-drug and drug-food interactions (DrugBank), chemical-target binding (BindingDB, ChEMBL), drug-likeness, and pharmacological classification (FDA classes, mechanisms of action, MeSH, ATC) on top of plain compound search. Surfaced 2026-08-02 Drug Repurposing and Discovery pass. Deferred as a same-data-source duplicate: PubChem is already catalogued (pubchem.md), so per the one-entry-per-tool rule this should be folded into that page as an alternative install path — not given its own entry. Do that once its license, tool list, and a copy-pasteable claude mcp add snippet are confirmed from the GitHub README. Same author as the catalogued openfda-mcp-server.md and protein-mcp-server.md, both of which have held up well. Most relevant to Chemistry / Drug Repurposing and Discovery.
• FDB MedProof MCP (First Databank) — announced GA 2026-04-03 as an MCP server purpose-built to ground AI agents in clinical drug knowledge (medication decision support). Surfaced 2026-08-02 Drug Repurposing and Discovery pass. Deferred: it is a commercial enterprise product sold on FDB’s existing drug-knowledge-base contracts, and no public enable path, pricing, or claude mcp add-compatible endpoint was located from a primary source — only trade-press coverage. Revisit if FDB publishes a connector listing or install documentation. Most relevant to Translational Medicine / Drug Repurposing and Discovery.
• Helix GenoSphere Connector — Anthropic Life Sciences directory connector (joined the 2026-06-30 Claude Science launch cohort with Inductive Bio / Revvity / Biomni Lab). Population genomics plus longitudinal clinical data surfaced to Claude via MCP. Deferred 2026-07-04 under the soft cap after cataloguing Inductive Bio + Revvity from the same cohort. Most relevant to Molecular and Cellular Biology / Translational Medicine / Integrative Structural and Computational Biology. Verify the directory enable path and access model (likely enterprise/account-gated like its cohort peers) before cataloguing; then create helix-genosphere.md.
• Biomni Lab Connector — the hosted, commercial “AI-native biology environment” connector from the same 2026-06-30 Claude Science cohort. Distinct from the catalogued open-source Stanford SNAP Lab biomni.md Skill (Apache-2.0, self-hosted Python agent). Deferred 2026-07-04 under the soft cap; when catalogued it needs a separate page (e.g., biomni-lab.md) with a clear “not the same as the open-source Biomni Skill” note, plus a verified enable path/access model. Tagged All (broad biomedical agent).
• SandboxAQ Claude Connector — Anthropic-official Claude.ai Connector (in the Connectors Directory, “Large quantitative models for scientific discovery”) integrating SandboxAQ’s LQMs via MCP (announced 2026-05-18). Deferred on the Drug Repurposing and Discovery directed pass (2026-06-28): the two drug-discovery models (AQPotency candidate prioritization, AQCell cellular-response/toxicity simulation) are still waitlist-gated “coming soon” per the launch press release; the only model live in Claude at announcement is AQCat (catalyst discovery, materials science). Access is via a waitlist (go.sandboxaq.com/LLMtoLQMWaitlist...), and pricing/install specifics for the directory connector are not published. Revisit once AQPotency/AQCell are GA in the Connectors Directory with a documented enable path — at which point it is a Drug Repurposing and Discovery entry; the catalyst-only surface would otherwise sit under General-Purpose Utilities.
• biorxiv@life-sciences plugin — anthropics/life-sciences marketplace entry currently DOA (mcp.deepsense.ai NXDOMAIN per upstream issue #42 / commit e96556b). Revisit when upstream restores the endpoint.
• clinical-trials@life-sciences plugin — same DOA status as biorxiv@life-sciences; distinct from both the adisinsight plugin (Springer Nature) and the clinicaltrials-gov-mcp community entry. Revisit when the endpoint is restored.
• Retrosynthesis MCP — partly resolved 2026-06-20: CovaSyn Chemistry MCP (covasyn.md) now exposes a retrosynthesis tool in its covaplatform suite (hosted, freemium), so a Claude-installable retrosynthesis path exists. Open-source single-tool wrappers remain absent: no discrete MCP/Skill for AiZynthFinder / ASKCOS / IBM RXN located on the 2026-05-25 → 2026-06-20 Chemistry passes (AiZynthFinder v4 ships a Flask/JSON API and ASKCOS a FastAPI service, neither MCP-wrapped). Reconfirmed absent on the 2026-07-04 Chemistry pass. Revisit if an OSS retrosynthesis wrapper ships.
• MoziChem-MCP (sinagilassi/mozichem-mcp, PyPI mozichem-mcp, MIT) — chemical-engineering thermodynamics MCP bundle: EOS Models server (Peng-Robinson/SRK/RK/vdW, fugacity) and Flash Calculations server (VLE, bubble/dew point), uvx-launched stdio. Surfaced 2026-06-13 Chemistry pass. Deferred as adjacent-domain (chemical-engineering thermodynamics rather than life-science chemistry); if catalogued, it belongs under General-Purpose Utilities. Revisit if life-science recipes start reaching for it.
• ZINC Database Skill (davila7/claude-code-templates, cli-tool/components/skills/scientific/zinc-database/SKILL.md) — installable scientific Claude Code skill over UCSF’s ZINC22 (~230M+ purchasable compounds; CartBlanche22 API). Partly resolved 2026-06-11: a ZINC15/ZINC22 query skill from SciAgent-Skills is now catalogued (zinc-database.md, Chemistry + Drug Repurposing and Discovery). The davila7/claude-code-templates variant remains a distinct alternative packaging; if its npx claude-code-templates@latest --skill path and SKILL.md tool list are later verified to followable-verbatim standard, fold them into zinc-database.md as an alternative install path rather than creating a second entry.
• MCP_Vina (shogo-d-nakamura/MCP_Vina) — AutoDock Vina MCP server with dock_molecule / get_available_targets tools; currently single-target (AKT1 only) with uv-based local install. Surfaced on the 2026-06-01 Chemistry pass but deferred because the target list is too narrow to justify a catalog entry today; revisit once additional targets or a docking-grid-builder workflow lands upstream.
• Ensembl MCP — munch-group variant (munch-group/ensembl-mcp) — still early-stage. The effieklimi/ensembl-mcp-server sibling was catalogued 2026-06-11 (ensembl.md), and the SciAgent Ensembl REST skill is folded into that same page as an alternative install path; one entry per data source already covers Ensembl, so the munch-group variant only warrants folding in as an alternative install path if it matures.
• UCSC Genome MCP (hlydecker/ucsc-genome-mcp) — 12 tools over the UCSC Genome Browser API. The UCSC REST API is now catalogued as a SciAgent skill (ucsc-genome-browser.md); fold this MCP in as an alternative install path when verified.
• NCBI Datasets MCP (Augmented-Nature/NCBI-Datasets-MCP-Server) — 31 tools over the NCBI Datasets API.
• OpenFDA MCP — Augmented-Nature variant (Augmented-Nature/OpenFDA-MCP-Server) — broader drug+device openFDA wrapper (10 active tools: 6 drug + 4 device incl. search_device_510k, search_device_classifications, search_device_adverse_events (MDR/MAUDE), search_device_recalls; food endpoints listed only as “Future Enhancements”, not implemented). Re-investigated 2026-06-21 on the Translational Medicine directed pass — the README’s tool list and git clone/npm install/npm run build steps are now confirmed, but still deferred: the README states only “available under standard licensing terms” (LICENSE file content not displayed) and provides no copy-pasteable claude_desktop_config.json / claude mcp add registration snippet (node path to the built server unstated), so it fails the followable-verbatim bar. The drug-endpoint subset is covered by the catalogued ythalorossy/openfda entry (openfda.md); the device superset (510k/MAUDE/classification/recall) remains the unique gap. Revisit once the explicit license and a literal registration snippet are confirmed from a primary source.
• Azure FHIR MCP (erikhoward/azure-fhir-mcp-server) — Azure Health Data Services FHIR adapter.
• AACT Clinical Trials MCP (navisbio) — read-only SQL over the CTTI AACT PostgreSQL mirror of ClinicalTrials.gov (70+ tables; list_tables / describe_table / read_query / append_insight); complementary to the v2-API clinicaltrials-gov-mcp entry. Re-investigated 2026-06-06 and still deferred: license reported inconsistently (MIT vs GPLv3 across directory listings) and the canonical repo/package path is ambiguous (navisbio/aact_mcp vs navisbio/mcp-server-aact, npm mcp-server-aact). Requires a free AACT account (DB_USER/DB_PASSWORD). Revisit once the LICENSE file and canonical package name are confirmed from the GitHub repo directly.
• Augmented-Nature ChEMBL-MCP-Server — 22-tool community alternative to the official Anthropic ChEMBL connector; useful for users without marketplace access. Deferred because the first-party connector now covers the same surface.
• easysolutions906 Healthcare MCP (@easysolutions906/mcp-healthcare) — 10-tool community bundle covering ICD-10, NPI, NDC, and DEA in a single server; complementary to the four discrete Anthropic Healthcare MCPs. Revisit on next Translational pass if users want a single-install alternative.
• IEDB MCP wrapper — partly resolved 2026-07-11: the bioSkills epitope-prediction skill (epitope-prediction.md) now provides a Claude-installable epitope workflow that wraps the IEDB REST API alongside BepiPred-3.0/DiscoTope-3.0/NetMHCpan/MHCflurry. A dedicated, standalone IEDB Query-API MCP server (exposing the ~2.2M epitopes / ~6.8M assays of the IQ-API directly as queryable tools, per the Nucleic Acids Research 2024 update) still does not exist and remains a natural MCP candidate. Revisit on next Immunology-focus pass.
• BCR/TCR repertoire MCP / Immcantation / AIRR-seq skill — partly resolved 2026-07-11: the bioSkills immcantation-analysis skill (immcantation-analysis.md) now wraps the Immcantation R suite (alakazam/shazam/scoper/dowser/tigger + IgBLAST/Change-O/IgPhyML) as a Claude-installable skill. bioSkills also ships mixcr-analysis, scirpy-analysis, vdjtools-analysis, repertoire-visualization, specificity-annotation (deferred — see the bioSkills batch entry). No dedicated MCP wrapper for nf-core/airrflow, immunarch, TRUST4, VDJdb, or McPAS-TCR located yet. Revisit on next Immunology-focus pass.
• Metagenomics / microbiome / AMR MCP — partly resolved 2026-07-11: the bioSkills kraken-classification (kraken-classification.md, Kraken2 + Bracken) and amr-detection (amr-detection.md, RGI/AMR++/deepARG/AMRFinderPlus) skills now provide Claude-installable metagenomics/AMR workflows; bioSkills also ships metaphlan-profiling, strain-tracking, functional-profiling, and a full microbiome/ (QIIME2/DADA2/phyloseq) category (deferred — see the bioSkills batch entry). No discrete MCP server for these CLI tools yet — florensiawidjaja/BioinfoMCP (arXiv 2510.02139) can auto-convert them but is a converter framework, not a discrete installable tool (out of scope, same rule as biocontext-ai/skill-to-mcp). Revisit if a pre-built metagenomics/AMR MCP ships.
• BoltzGen Claude skill — the all-atom binder/nanobody/antibody design model HannesStark/boltzgen is itself out of scope (model weights + PyPI package, a research artifact). A third-party “BoltzGen Claude Code skill” is listed on mcpmarket.com but its upstream source repo / SKILL.md could not be located or verified on the 2026-06-13 Immunology pass (mcpmarket returned HTTP 429). Revisit if a verifiable Claude-skill wrapper with a primary source surfaces.
• MDAnalysis (K-Dense Skill) — standalone trajectory-analysis sibling to the molecular-dynamics skill; may warrant its own entry only if the K-Dense collection ever ships a distinct mdanalysis SKILL.md directory (it currently does not — the trajectory analysis lives inside molecular-dynamics).
• Cryo-EM MCP / Skill — no Claude-installable wrapper for RELION / cryoSPARC / CTFFIND located on the 2026-05-27, 2026-06-03, 2026-06-13, 2026-06-20, 2026-06-27, 2026-07-04, 2026-07-25, or 2026-08-01 Structural passes; cryoSPARC v5.0 (2026-01-27) and RELION remain CLI/GUI-only with no MCP/Skill wrapper (GROMACS’s built-in correlation-based density fitting is the closest MD-side path but is not a standalone tool). The 2026-08-01 pass confirmed the unwrapped building block is cryoem-uoft/cryosparc-tools (pip-installable Python scripting API, usable outside the cryoSPARC environment) — an MCP wrapper over it would be the natural entry, and the right shape is a job-queue/Slurm-submitting server rather than GUI automation, since both packages are HPC-bound. Name-collision warning for future searches: z80dev/cryo-mcp is a blockchain data server, unrelated to microscopy. Partial coverage exists via the catalogued structure-validation.md skill (cryo-EM global-vs-local resolution, FSC 0.143 half-map vs 0.5 map-model) and emdb-database.md (map retrieval). Revisit on next Integrative Structural and Computational Biology pass.
• trRosettaRNA MCP / RNA 3D structure prediction — surfaced on the 2026-08-08 Integrative Structural and Computational Biology pass via a LobeHub listing for biomolecular-design-nexus/trrosettarna_mcp (advertised: MSA-based RNA 3D prediction, dot-bracket/CT/SPOT-RNA secondary-structure constraints, geometry-only contact/distance mode, documented claude mcp add path). Deferred — unverifiable: https://api.github.com/repos/biomolecular-design-nexus/trrosettarna_mcp returns 404 (private or moved), so license, tool list and install snippet have no primary source. Same org as the deferred boltz_mcp. The underlying method (trRosettaRNA2, Nat Mach Intell 2026) is a research artifact and out of scope on its own. RNA 3D structure prediction remains a complete catalog gap — the newly-catalogued rna-structure/ skills (covariation-analysis.md, ncrna-search.md, structure-probing.md) and viennarna-structure-prediction.md are all secondary-structure-level. Revisit if the repo becomes public or another RNA-3D wrapper ships. Most relevant to Integrative Structural and Computational Biology.
• FoldMason / standalone TM-align MCP — no discrete MCP or Skill wrapping the multiple-structure-alignment tools steineggerlab/foldmason (Gilchrist, Mirdita & Steinegger, Science 2026 — progressive MSTA over Foldseek + TM-align, two orders of magnitude faster than prior MSTA methods) or rcedgar/reseek located on the 2026-08-01 Structural pass. TM-align/jFATCAT and Foldseek remain reachable to Claude only inside the catalogued protein-mcp-server.md (protein_compare_structures, protein_find_similar) and foldseek-structural-search.md; there is no pairwise/multiple-alignment server of their own. Partly resolved 2026-08-15: the bioSkills structural-alignment skill (structural-alignment.md) is now catalogued and drives FoldMason, TM-align, US-align, Foldseek-Multimer, DALI, FATCAT and CE as local CLIs with score-interpretation bands, so a Claude user has a followable path to structural MSA and pairwise superposition. What still does not exist is a discrete MCP server for any of them — the remaining gap is an MCP wrapper, not a usable path. Revisit if such a server ships. Most relevant to Integrative Structural and Computational Biology.
• chimerax-alphafold-mcp (GDAmitha/chimerax-alphafold-mcp) — sibling to the catalogued chimerax-mcp.md that additionally wires AlphaFold/ESMFold prediction directly into ChimeraX (molecular visualization + structural analysis + arbitrary ChimeraX Python + prediction). Surfaced on the 2026-07-25 Integrative Structural and Computational Biology directed pass; the mahynotch/chimerax-mcp variant was catalogued this run as the cleaner ChimeraX driver. Fold this in as an alternative install path (or a distinct entry if the prediction surface is materially different) once its license and a copy-pasteable registration snippet are confirmed from a primary source.
• gmx-vmd-mcp (egtai/gmx-vmd-mcp) — GROMACS + VMD integration MCP (workflow create/manage, simulation prep, RMSD/RMSF trajectory analysis, VMD visualization templates) over an HTTP microservice (localhost:8000); prereqs GROMACS + VMD in PATH. Surfaced on the 2026-06-20 Integrative Structural and Computational Biology directed pass (GROMACS molecular dynamics MCP server query) but deferred: the README’s clone URL is a yourusername/mcp-gmx-vmd placeholder, the discrete tool names are not enumerated (only /gmx-vmd://workflow/... endpoint patterns), and the LICENSE was not confirmed — not followable-verbatim. The standalone GROMACS MD-runner gap is now closed by the Docker-based MacromNex/gromacs_mcp (catalogued 2026-07-11, gromacs-mcp.md); gmx-vmd-mcp would only add distinct value for its VMD-visualization tools. Revisit once the canonical repo path, a copy-pasteable registration snippet, and an explicit tool list are confirmed from a primary source. (OpenMM MCP, the cleaner MD-runner candidate from the same pass, was also catalogued — openmm-mcp.md.)
• RFdiffusion / ProteinMPNN Claude Skill — protein-design model wrappers; standalone Claude Skill / MCP not located on the 2026-05-27 or 2026-06-03 sweeps. Both are bundled inside the longevity-genie/protein_hunter_mcp server (which also includes Boltz / Chai-lab / PyRosetta / LigandMPNN) — that server is a viable structural-biology entry once its install path is verified outside the longevity-genie organisation.
• Boltz / Protein Hunter MCP (longevity-genie/protein_hunter_mcp, biomolecular-design-nexus/boltz_mcp) — Boltz2 structure / affinity prediction wrappers. Surfaced on the 2026-06-03 Structural pass but deferred — Boltz model weights are MIT but the protein_hunter_mcp install path requires uv + CUDA GPU and the upstream README does not yet publish a copy-pasteable claude mcp add snippet.
• mne-neurophysiology-analysis — listed on mcpmarket.com/tools/skills/mne-neurophysiology-analysis but no verifiable upstream GitHub source / SKILL.md found via 2026-06-05 WebSearch (neither in K-Dense-AI/scientific-agent-skills nor in mne-tools org). Investigated on the 2026-06-05 Neuroscience pass and deferred until a primary source surfaces. Note: MNE-Python is now covered by the catalogued NeuroClaw mne-eeg-tool skill (mne-eeg-tool.md), so the gap is largely closed; if this specific mcpmarket listing resolves to a real repo, fold it in or add as a discrete entry as appropriate.
• HED-MCP (neuromechanist/hed-mcp, BSD-3-Clause) — MCP server automating HED (Hierarchical Event Descriptor) sidecar generation/validation for BIDS event files: analyze_event_columns, generate_hed_sidecar, validate_hed_string, validate_hed_file, list_hed_schemas. Surfaced on the 2026-06-28 Neuroscience directed pass. Deferred — not installable today: the README states it is “in active development … implementation coming soon” and the PyPI package (hed-mcp) is not yet published, so the documented pip install hed-mcp path does not resolve. Re-checked 2026-08-02: https://pypi.org/pypi/hed-mcp/json still returns 404 — no release yet. Revisit once hed-mcp ships to PyPI and the server is runnable. A natural complement to the catalogued bids skill.
• brain-bbqs/NeuroMCP (io.github.brain-bbqs/neuro-mcp in the MCP Registry, v1.1.1) — “Neuroscience MCP server (Python) for the BBQS at McGovern Institute of Brain Research” (NIH BRAIN Initiative Brain Behavior Quantification & Synchronization program). Surfaced on the 2026-06-28 Neuroscience directed pass via the official MCP Registry. Deferred: the GitHub repo brain-bbqs/NeuroMCP returns 404 (private or moved), so the tool list, license, and a copy-pasteable install/registration snippet could not be verified from a primary source. Revisit once the repo is public and the install path is confirmed.
• sensein/brainkb_mcp (org.brainkb/brainkb v0.1.0, homepage https://mcp.brainkb.org) — MCP server over BrainKB, the Sense Lab / BICAN neuroscience knowledge graph, exposing its query_service API: session/auth tools (brainkb_login, brainkb_globus_login, brainkb_use_token, brainkb_create_token, brainkb_list_tokens, brainkb_revoke_token, brainkb_whoami, brainkb_logout), space management (brainkb_list_spaces, brainkb_create_space, brainkb_set_space_visibility, brainkb_add_space_member), plus knowledge-graph ingest, search, provenance query and ingest-job management. Auth is mandatory — a BRAINKB_TOKEN PAT or Globus/ORCID/GitHub OAuth; there is no anonymous read. Surfaced on the 2026-08-16 Neuroscience directed pass. Deferred on two blockers: (1) the GitHub license API returns null — the repo ships no LICENSE, so it fails the license-clear bar; (2) the README labels the remote URL https://mcp.brainkb.org/mcp as “(future)”, so the one-line hosted install path is not confirmed live. Would be a genuinely novel entry — the catalog has no neuroscience knowledge-graph surface at all. Revisit once a license lands and the remote endpoint is confirmed serving.
• ABC Atlas MCP or Skill — standing gap, no candidate found (2026-08-16 Neuroscience pass). The catalogued allenbrain.md / allenbrain-mcp wraps the older Allen Brain Atlas RMA API; nothing wraps the Allen Brain Cell (ABC) Atlas abc_atlas_access stack (AbcProjectCache, public S3 bucket arn:aws:s3:::allen-brain-cell-atlas, no login required — whole mouse brain transcriptomic atlas, Consensus Whole Mouse Brain, Aging Mouse, MERFISH-to-CCF mapped coordinates, v1.3.1 / release 20260331). No MCP server and no Skill located. Because the data is a keyless public S3 bucket with a documented Python cache object, a Skill wrapping abc_atlas_access is the lower-friction shape than a full MCP server. Also re-confirmed absent this pass: any connectome-specific MCP, and any calcium-imaging pipeline (Suite2p / CaImAn) MCP or Skill.
• bendichter/dandi-query-mcp — TypeScript MCP server advertised as a “DANDI Archive query server” on Glama and Lobehub, but the published tool list is the boilerplate create_note / summarize_notes MCP template, not a DANDI-API wrapper; investigated on the 2026-06-05 Neuroscience pass and not added. DANDI semantic search and dandiset_info / dandiset_assets / dandi_list_neurodata_types are already covered by neurosift (Neurosift Tools MCP). Revisit only if upstream replaces the boilerplate tools with real DANDI-API tools.
• NWB conversion — still absent; NWB reading is now closed (updated 2026-08-16 Neuroscience pass). Resolved for reading: bjhardcastle/nwb-mcp-server (MIT, PyPI 0.1.9) was catalogued 2026-08-16 as nwb-mcp-server.md — read-only SQL query over local/S3/DANDI NWB files. Still open: nothing wraps catalystneuro/neuroconv, NWB GUIDE, or NWB Inspector, so conversion-to-NWB and NWB validation remain unreachable. Original note follows. No MCP server or Skill wraps catalystneuro/neuroconv or NWB Inspector. CatalystNeuro publishes MCP servers but only general-purpose ones (e.g. a “Read Images” vision server). NWB/DANDI remain reachable to Claude only through the catalogued neurosift.md and spikelab.md (the latter exposes MCP tools for all library methods via pip install "spikelab[s3,ml,mcp]"). Conversion-to-NWB is the clean unfilled gap; revisit if a wrapper ships. Name-collision warning for future searches: the nemad server in pathintegral-institute/mcp.science is not neuroscience — NEMAD is a magnetic/thermoelectric/superconductor materials database (nemad_search, nemad_formula_search, nemad_read_results, NEMAD_API_KEY). If ever catalogued it belongs under General-Purpose Utilities alongside materials-project-mcp.md, and it largely duplicates that page’s audience.

User requests (open)

User requests (blocked)

• [#98 @QuantaAIBot 2026-08-19] queue: catalog

  request=new-tool

  name=”FDA 510(k) product-code CSV exporter”

  url=”https://github.com/QuantaAIBot/fda-510k-product-code-export”

  subject_area=”Translational Medicine”

  author=@QuantaAIBot

  issue=98

• [#100 @vladimir-mitev 2026-08-26] queue: catalog

  request=new-tool

  name=”Folklore Clinical Variant Interpretation MCP”

  url=”https://github.com/helena-bioinformatics/folklore-mcp”

  subject_area=”Translational Medicine”

  author=@vladimir-mitev

  issue=100

• [#101 @helena-good-outreach 2026-08-31] queue: catalog

  request=new-tool

  name=”Noodle Biomedical Literature Discovery MCP”

  url=”https://github.com/helena-bioinformatics/noodle-mcp”

  subject_area=”Translational Medicine”

  author=@helena-good-outreach

  issue=101

User requests (closed this run)

None.