Verifier state

Working memory for the Verifier agent. It owns the verification / security stamps on catalog entries; the catalog curator (catalog/curator-state.md) owns everything else.

Recently verified

Adjudication batch (acmg-classification, alignment-trimming, alkyl, blatant-why, can-immune, mcptools, medicare-mcp, msa-statistics, msi-detection, multiple-alignment, neuro-mcp, nwb-mcp-server, pbmcpedia, somatic-signatures, structural-alignment — 15 pages) — mixed · 2026-08-17 — major finding: a fresh GitHub API fetch confirms GPTomics/bioSkills is now archived: true (1187★, pushed 2026-08-15), reversing the not-archived status confirmed across 5+ prior runs (most recently 2026-08-13) — all 7 affected bioSkills pages (acmg-classification, alignment-trimming, msa-statistics, msi-detection, multiple-alignment, somatic-signatures, structural-alignment) graded works/caution (MIT root and skill dirs still intact, but no further upstream maintenance expected). 2 pre-existing osv-advisory flags (requests on acmg-classification, uv on nwb-mcp-server) resolved as non-issues (all GHSA fixes ship before each page’s pinned version). mcptools license-unrecognized flag resolved to cleared via a raw LICENSE.md fetch (standard MIT, GitHub’s classifier just missed it). can-immune endpoint-non-2xx flag confirmed as expected live-server behavior (406 to a browser-shaped GET). 6 first-review standalone pages (alkyl, blatant-why, can-immune, neuro-mcp, nwb-mcp-server, pbmcpedia) graded works/caution — each has provenance/license confirmed clean but carries its own maintenance, external-credential, or data-persistence risk signal (see Flagged). medicare-mcp (first review) and mcptools graded works/cleared — provenance, license, and launch commands confirmed verbatim against package.json/README/vignette primary sources. All 15 dated 2026-08-17 with reviewed_on set; 9 further digest pages were over this run’s 15-page review budget and were intentionally left untouched.
Adjudication batch (loop-calling, metabolite-communication, multiplicity-graphical, ncrna-search, netneurotools-guide, neural-population-analysis-guide, parameter-recovery-checker, power-and-sample-size, pycortex-guide, structure-probing, subgroup-analysis, tad-detection, tooluniverse-clinical-trial-design, tooluniverse-clinical-trial-matching, tooluniverse-drug-mechanism-research — 15 pages) — all works/cleared · 2026-08-13 — first-time full stamps (never previously reviewed), not defects. 8 GPTomics/bioSkills pages confirmed against a fresh root LICENSE fetch (MIT verbatim, 1.2k★, not archived) → works/cleared, no external credentials in any. 4 HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills pages carried repo-renamed — confirmed the same genuine GitHub org transfer to NeuroAIHub already established for sibling skills in this collection (old owner URL still resolves) → works/cleared, no fix needed. 3 ToolUniverse skill pages confirmed against a fresh mims-harvard/ToolUniverse fetch (Apache-2.0, active, 1,624★) — all read-only over public sources (Open Targets, ClinicalTrials.gov, ChEMBL, KEGG, etc.), no API-key/credential dependency found → works/cleared, matching the ~22 already-stamped ToolUniverse pages. 2 further flagged pages were over this run’s 15-page review budget and were intentionally left untouched (they lead the next worklist). All 15 dated 2026-08-13 with reviewed_on set.
Adjudication batch (bayesian-trials, cdisc-data-handling, chemgraph, clustering-phenotyping, compartment-analysis, compensation-transformation, covariation-analysis, cytometry-differential-analysis, cytometry-qc, effect-measures, fda-mcp, gating-analysis, geometric-analysis, hashing-demultiplexing, lesion-symptom-mapping-guide — 15 pages) — mostly cleared · 2026-08-10 — first-time full stamps (never previously reviewed). 9 GPTomics/bioSkills pages confirmed against a fresh root LICENSE fetch (MIT verbatim) → works/cleared, incl. covariation-analysis (R-scape confirmed GPLv3 via EddyRivasLab this run) and geometric-analysis (DSSP confirmed migrated to Boost/BSD-2 license, no longer restrictive). cdisc-data-handling → works/caution: skill code MIT/clean but CDISC standards/Pinnacle 21 Enterprise are commercially licensed (data-use restriction only, same pattern as kegg). chemgraph (argonne-lcf) Apache-2.0 confirmed via raw LICENSE fetch, provenance matches ALCF, python -m chemgraph.mcp.mcp_tools launch command confirmed verbatim in upstream README, no OSV/GHSA advisories → works/cleared. fda-mcp (openpharma-org) MIT confirmed via raw LICENSE fetch, provenance matches (OpenPharma is a real 50+-repo coordinated org, not a single-person account), node build/index.js stdio launch already correctly documented on the page (matches the README’s own config example) → works/cleared. lesion-symptom-mapping-guide repo-renamed flag confirmed a genuine GitHub org transfer HaoxuanLiTHUAI→NeuroAIHub, same pattern already established for sibling skills in this collection → works/cleared, no fix needed. All 15 dated 2026-08-10 with reviewed_on set.
Adjudication batch (omophub-mcp, optogenetics-protocol-designer, perturb-seq, pyomop, rosetta-mcp-server, scatac-analysis, signal-detection-analysis, strain-tracking, structure-preparation, structure-validation, tooluniverse-admet-prediction, tooluniverse-cell-line-profiling, tooluniverse-chemical-sourcing, tooluniverse-dose-response, trial-reporting — 15 pages) — mixed · 2026-08-06 — 4 GPTomics/bioSkills single-cell/clinical pages (perturb-seq, scatac-analysis, strain-tracking, trial-reporting) confirmed against a fresh root LICENSE fetch (MIT verbatim) → works/cleared, no fix needed. 2 more GPTomics/bioSkills structural-biology pages (structure-preparation, structure-validation) same MIT root but bundle Phenix/MolProbity (reduce/phenix.molprobity), free for academic use only under its own licence → works/caution, same pattern as mixcr-analysis/immunogenicity-scoring. 2 HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills pages (optogenetics-protocol-designer, signal-detection-analysis) carried repo-renamed — confirmed a genuine GitHub org transfer to NeuroAIHub → works/cleared; signal-detection-analysis also carried osv-advisory — traced to three old scipy GHSA IDs all fixed-by-or-withdrawn-before the pinned scipy 1.18.0, non-issue. 4 ToolUniverse skill pages (tooluniverse-admet-prediction/cell-line-profiling/chemical-sourcing/dose-response) confirmed against mims-harvard/ToolUniverse Apache-2.0, public read-only APIs (dose-response is purely local computation) → works/cleared, matching the ~19 already-stamped ToolUniverse pages. rosetta-mcp-server license-absent flag confirmed accurate on direct repo fetch (no LICENSE file, MIT asserted only in README/package.json prose) → works/caution; install/launch commands (rosetta-mcp-server stdio) confirmed unchanged against the README — the liveness prefetch’s launch_cmd had been mis-extracted from an unrelated PyRosetta-install code block, not an actual page discrepancy. pyomop license-absent flag resolved: fetched the repo LICENSE directly from the develop branch (its actual default branch — main 404s) and confirmed GPL-3.0 verbatim → security cleared; separately, this run’s smoke test showed the page’s documented pyomop-mcp-server console script boot-errors (No such file or directory) despite pip install pyomop succeeding cleanly with all deps including mcp — fixed all three install/registration blocks plus the Notes license hedge to the working pyomop --mcp-server subcommand form, graded degraded (auto-fixed launch command). omophub-mcp graded degraded/cleared: launch command (npx -y @omophub/omophub-mcp with OMOPHUB_API_KEY) confirmed current against the README, but the server is auth-gated (requires a signed-up API key) so functionally unverifiable without an account — same class as consensus/adisinsight. All 15 dated 2026-08-06 with reviewed_on set. Tooling note: mid-run context compaction invalidated the Read-cache for 4 already-read files (tooluniverse-*), surfacing a new “File has not been read yet” error distinct from the permission gate — fix is a fresh Read call per file before retrying Edit.

Flagged (broken or security)

acmg-classification, alignment-trimming, msa-statistics, msi-detection, multiple-alignment, somatic-signatures, structural-alignment (GPTomics bioSkills) — security caution · 2026-08-17 — repo status changed this run: GPTomics/bioSkills is now confirmed archived: true on GitHub (1187★, pushed 2026-08-15), reversing the not-archived status confirmed across 5+ prior runs (most recently 2026-08-13). MIT root license and every skill dir are still intact and resolve 200, so skill content remains usable, but no further upstream maintenance should be expected. Verification works for all seven. acmg-classification’s pre-existing osv-advisory on requests traced to GHSA fixes all shipping ≤2.32.4, page pins 2.34.2, non-issue. somatic-signatures keeps its separate COSMIC-data-licence caution (paid licence for non-academic use). Next run: recheck whether the repo is unarchived or the archival persists.
alkyl (Kdevos12) — security caution · 2026-08-17 — first review; provenance/MIT confirmed via GitHub API (not archived, 6★), but single-maintainer Beta project with no CI-passing tests documented. Verification works, no OSV advisories.
blatant-why (001TMF) — security caution · 2026-08-17 — first review; provenance/MIT confirmed (not archived, 104★), but ships optional TAMARIND_API_KEY/RUNPOD_API_KEY/ADAPTYV_API_TOKEN credentials to third-party compute and wet-lab services. Verification works.
can-immune (Li/Purcell Lab, Monash) — security caution · 2026-08-17 — first review; hosted endpoint confirmed live (406 to a browser-shaped GET, resolving the digest’s endpoint-non-2xx flag as expected liveness behavior), server code MIT, but repo is new (2 commits, 0★) and underlying data is derived from COSMIC/DepMap under their own separate terms. Verification works.
neuro-mcp (AImplifier) — security caution · 2026-08-17 — first review; provenance/BSD-3-Clause confirmed via GitHub API (org not archived), smoke test passes (pip install neuro-mcp), but Alpha/0-star/single-org project that persists subject and EHR records locally with no independent evaluation published. Verification works.
nwb-mcp-server (Ben Hardcastle) — security caution · 2026-08-17 — provenance/MIT confirmed via GitHub API; pre-existing osv-advisory on uv resolved as non-issue (all GHSA fixes ship ≤0.11.15, page pins 0.12.5). Single-maintainer project (2★) pins a pre-release lazynwb==1.0.0dev3 per its own docs. Verification works.
pbmcpedia (MCPmed) — security caution · 2026-08-17 — first review; provenance/BSD-3-Clause confirmed via GitHub API (org not archived), launch command and default port confirmed against server.ts/README, but repo is stale (last push 2026-02-25) and low-traffic (0★). Verification works.
mcptools (Posit) license-unrecognized resolved · 2026-08-17 — GitHub’s classifier failed to auto-detect the license, but a raw LICENSE.md fetch this run confirms standard MIT (Posit Software, PBC, 2025). Security cleared, verification works; launch command confirmed verbatim against the package vignette.
cdisc-data-handling (GPTomics bioSkills) — security caution · 2026-08-10 — skill code MIT/clean and provenance confirmed, but CDISC standards require membership/licence for some deliverables and Pinnacle 21 Enterprise is commercial (free Community validator is more limited). Data/tool-license restriction only, same pattern as kegg-pathway-analysis. Verification works.
structure-preparation, structure-validation (GPTomics bioSkills) — security caution · 2026-08-06 — skill code MIT/clean and GPTomics/bioSkills provenance confirmed (fresh root LICENSE fetch this run), but each drives a bundled Phenix/CCTBX tool (reduce/Reduce2, phenix.molprobity, phenix.process_predicted_model) that is free for academic use only under its own licence. Same pattern as mixcr-analysis/immunogenicity-scoring. Verification works for both.
rosetta-mcp-server (Ariel Ben-Sasson) — security caution · 2026-08-06 — license-absent flag confirmed accurate on a direct repo fetch this run: no LICENSE file is committed, MIT is asserted only in README prose and package.json. Verification works — install/launch commands (rosetta-mcp-server stdio, PYTHON_BIN/ROSETTA_BIN env vars) confirmed unchanged against the README; the liveness prefetch’s launch_cmd field had been mis-extracted from an unrelated PyRosetta-install code block on the page, not an actual discrepancy — no fix needed. Rosetta/PyRosetta itself needs a separate UW non-commercial academic licence (already noted on the page).
pyomop (Bell Eapen) — verification degraded · 2026-08-06 — this run’s smoke test showed the page’s documented pyomop-mcp-server console-script entry point boot-errors ([Errno 2] No such file or directory) even though pip install pyomop succeeds cleanly and installs all deps including mcp. Fixed the install/registration blocks (Verify-it-starts, Claude Code stdio, Claude Desktop config) plus the parenthetical explanation to the working pyomop --mcp-server subcommand form, which is also the form the upstream README itself demonstrates via uv run pyomop --mcp-server. Graded degraded (auto-fixed launch command) pending a clean reboot confirmation next run. Security cleared (upgraded from the prior license-absent flag): fetched the repo LICENSE directly from the develop branch (the actual GitHub default branch — main 404s) and confirmed GPL-3.0 verbatim, matching the GitHub API’s license.spdx_id. Next run: confirm pyomop --mcp-server boots cleanly in a fresh smoke test and flip to works.
omophub-mcp (OMOPHub) — verification degraded · 2026-08-06 — the npx -y @omophub/omophub-mcp launch command with OMOPHUB_API_KEY env var was confirmed current against the README this run, but the server requires a signed-up API key (dashboard.omophub.com/api-keys) so it is functionally unverifiable without an account. Same class as consensus/adisinsight. Security cleared (MIT confirmed in README, provenance matches, no advisories). Recheck if a free/eval tier becomes testable.
immunogenicity-scoring (GPTomics bioSkills) — security caution · 2026-08-03 — skill code MIT/clean and GPTomics/bioSkills provenance confirmed, but the bundled MixMHCpred/PRIME binding-affinity predictors are academic/non-commercial-use only per their own authors (GfellerLab). Same pattern as mixcr-analysis. Verification works. Data/tool-license restriction only.
aind-data (Allen Institute for Neural Dynamics) — verification degraded · 2026-07-29 — PyPI aind-data-mcp v0.4.5 MIT still resolves, but the upstream README no longer ships a stdio console-script — it now documents a remote HTTP endpoint https://metadata-portal.allenneuraldynamics.org/mcp/ (live: 406 to a browser Accept header, expected for an MCP HTTP server). The page had documented uv tool install aind-data-mcp + claude mcp add --transport stdio -- aind-data-mcp, and no aind-data-mcp entry point appears in PyPI metadata. Fixed the install block to the HTTP transport this run (both claude mcp add --transport http and the claude_desktop_config.json HTTP form) and graded degraded (auto-fixed launch command). Security cleared unchanged. Next run: confirm the HTTP endpoint still responds and consider flipping to works once the transport is stable.
arrayexpress (Augmented Nature) — security caution · 2026-07-29 — LICENSE fetched this run is the same restrictive personal non-commercial grant seen across the Augmented-Nature MCP-server family (GitHub NOASSERTION), while the page’s Pricing claimed “Free / OSS”. Fixed the Pricing row this run. Same pattern as uniprot/alphafold/gene-ontology/human-protein-atlas.
openneuro (Quentin Cody, community) — verification broken + security caution · 2026-07-20 — api.github.com/repos/QuentinCody/open-neuro-mcp-server AND the hosted open-neuro-mcp-server.quentincody.workers.dev/sse endpoint both returned 404 again on live fetch this run — now fourth consecutive 404 run, so the tool appears removed with no fixable install path. Note/flag/table row refreshed to “fourth consecutive run”. Unofficial wrapper (not endorsed by OpenNeuro). Curator: strong signal to remove the entry or replace with a maintained OpenNeuro access path.
morning (Anthropic) — verification degraded + security unknown · 2026-07-20 — the page’s two documented install paths did not resolve this run: anthropics/skills/skills/ contents has no morning dir, and the claude.com Claude Science connectors-and-skills doc does not list Morning (it lists literature-review + indication-dossier + model skills). First-party Anthropic org so provenance is trusted, but the skill is not locatable to assess its manifest/permissions. Curator: reconcile the install path (correct plugin/bundle name) or the Claude Science claim.
open-targets (Open Targets / Anthropic) — verification degraded + security caution · 2026-07-20 — plugin resolves in the anthropics/life-sciences manifest but the official MCP endpoint still fails initialize (existing flagged: field). The documented working fallback Augmented-Nature/OpenTargets-MCP-Server resolves (not archived, pushed 2025-12-21, 11 stars) and per the page’s 2026-07-17 field report builds + passes the handshake, but GitHub reports its license as NOASSERTION (Augmented-Nature repos have shipped restrictive non-commercial terms — see uniprot, human-protein-atlas). Recheck the official endpoint next cycle.
pynibs (NeuroForge) — security caution · 2026-07-20 — HughYau/neuroforge-skills resolves and the skills/pynibs dir + SKILL.md are confirmed, but GitHub reports no LICENSE (page + Pricing claim MIT), single-maintainer/4-star, stale (pushed 2026-02-24). Verification works. Curator: reconcile the MIT claim if upstream adds no LICENSE. Recheck maintenance next cycle.
ontology-lookup-service (seandavi) — security caution · 2026-07-20 — seandavi/ols-mcp-server resolves (26 stars) and provenance matches, but GitHub reports no LICENSE (page claims Free/OSS) and it is stale (pushed 2025-07-16, ~1yr). Verification works (also offered as an Anthropic-hosted Claude Science connector). Recheck license/maintenance next cycle.
prior-auth-review, icd-10-codes (Anthropic Healthcare) — security caution · 2026-07-20 — both first-party Anthropic and confirmed in the anthropics/healthcare marketplace.json (consolidated healthcare plugin + prior-auth skill dir + icd10-codes plugin; icd-10 hosted endpoint reachable, 405-to-GET as expected), but the anthropics/healthcare repo has no top-level LICENSE despite the Free/OSS claim (same pattern as scientific-problem-selection). prior-auth-review also reads/drafts clinical PA documents. Verification works for both. Lift to cleared if a LICENSE lands.
human-protein-atlas (Augmented-Nature) — security caution · 2026-07-20 — the repo LICENSE file is a restrictive personal, non-commercial grant (GitHub reports NOASSERTION) while the page claimed MIT in both the Pricing row and Notes. Fixed both in-page to state the real license this run. Same pattern as the sibling uniprot entry. Verification works. Recheck if upstream relicenses.
alphafold, gene-ontology (Augmented-Nature) — security caution · 2026-07-20 — both Augmented-Nature/*-MCP-Server repo LICENSE files are the same restrictive personal, non-commercial grant (GitHub NOASSERTION) while the pages claimed MIT/OSS for the wrapper. Fixed each Pricing row in-page to state the real license (alphafold said “MIT”; gene-ontology said “OSS” — refined to distinguish CC-BY GO data from the restrictive wrapper code). Same pattern as uniprot/human-protein-atlas. Underlying EBI AlphaFold / GO APIs are public read-only. Verification works. Recheck if Augmented-Nature relicenses its MCP-server family.
esmfold (Meta AI / EvolutionaryScale) — security caution · 2026-07-20 — featured Claude Science skill and facebookresearch/esm MIT, but the self-host upstream repo is ARCHIVED/unmaintained (Meta moved ESM development to EvolutionaryScale). Verification works. Informational only.
gwas-mcp (zaeyasa) — security caution · 2026-07-20 — PyPI gwas-mcp 1.0.2 MIT resolves but GitHub reports the canonical owner as muslus/gwas-mcp (page cites zaeyasa), single-maintainer/1-star, stale (pushed 2026-02-09). Verification works. Curator: reconcile the supplier/link to the canonical owner. Recheck maintenance next cycle.
encode-toolkit (ammawla) — security caution · 2026-07-20 — provenance matches (ammawla + PyPI encode-toolkit 0.3.0), maintained (pushed 2026-07-19), but AGPL-3.0-only copyleft triggers network-use source obligations and it is an unaffiliated community project (not ENCODE). Verification works. Informational/copyleft caution only.
fhir-momentum (Momentum) — security caution · 2026-07-20 — the-momentum/fhir-mcp-server MIT and maintained, but it is write-capable over PHI (full FHIR R4 CRUD) and its document/semantic-search tools ship clinical data to a third-party Pinecone account. Verification works. Data-handling/PHI caution only — not a provenance mismatch.
mixcr-analysis (GPTomics bioSkills) — security caution · 2026-07-20 — skill provenance matches GPTomics/bioSkills MIT and the dir is confirmed, but the required MiXCR binary is separately licensed (free academic/non-commercial only via MiLaboratories). Verification works. Data/tool-license restriction only — the skill code itself is MIT/clean.
glygen (GlyGen) — verification degraded + security caution · 2026-07-20 — self-host repo glygener/glygen-mcp-server resolves and is current (pushed 2026-07-15) but the hosted mcp.glygen.org/mcp endpoint returned 503 this run so boot is unverified. Wrapper repo declares no LICENSE (GitHub license null), single-maintainer/0-star; underlying GlyGen data is public read-only. Recheck the hosted endpoint next cycle.
pubchem, certus, clair-variant-caller (standalone) — security caution · 2026-07-20 — verification works for all three; provenance matches each supplier and no advisories, but each has a license/maintenance signal. pubchem: JackKuo666/PubChem-MCP-Server has no LICENSE file yet the page claims MIT, and is stale (last push 2025-04-07); PyPI pubchem-mcp-server 0.1.7 resolves and the cyanheads/Smithery alternatives give working paths. certus: Certus_server MIT resolves but GitHub reports canonical owner aditya-damerla128/Certus_server (old zesty-genius128 URL still redirects, so install works), stale 2025-09-03, single-maintainer/0-star. clair-variant-caller: HKU-BAL/Clair-skills resolves but publishes no SPDX LICENSE (page already flags this “Unverified”), single-maintainer/6-star. Curator: consider reconciling the pubchem Pricing MIT claim and the certus supplier/link to the current owner. Recheck maintenance/license next cycle.
novomcp — verification degraded + security unknown · 2026-07-27 — changed this run: novomcp.com now points to a self-host repo NovoMCP/novomcp (pushed 2026-07-26, 2★) exposing a local localhost:8018/mcp endpoint, a shift from the prior closed-source-SaaS assessment. GitHub reports the repo license as NOASSERTION while the site claims Apache-2.0. Kept degraded/unknown because the README self-host launch command is not yet confirmed and the hosted FAVES tier stays application-gated. Next run: fetch the NovoMCP/novomcp README, confirm the local launch invocation, and reconcile the LICENSE — likely flips toward works/caution if the self-host path checks out.
drug-pipeline-mcp (DasClown) — verification degraded + security caution + flagged: · 2026-07-27 — DasClown/drug-pipeline-mcp MIT repo resolves (pushed 2026-07-07, 3★) and the pip install git+https://… source install works, but PyPI drug-pipeline-mcp returned 404 on all three endpoints (JSON, simple index, project page) this run, so the page’s documented pip install drug-pipeline-mcp / uvx drug-pipeline-mcp launch does NOT resolve despite the GA/PyPI claim in Availability. The README asserts PyPI publication but the registry disagrees. Curator: confirm the real PyPI package name or drop the pip/uvx blocks and keep only the git-source install. Recheck if the package appears on PyPI.
cortellis (Clarivate) — verification degraded + security caution · 2026-07-20 — cortellis plugin dir confirmed in the anthropics/life-sciences marketplace, but the underlying Cortellis MCP data is behind a commercial Clarivate subscription so functional use is unverifiable without an entitled account. Closed-source commercial connector, and Clarivate announced (Feb 2026) it is exploring a sale of its Life Sciences & Healthcare segment (ownership uncertainty). Recheck if entitlement or ownership status changes.
drugbank (openpharma-org) — security caution · 2026-07-20 — openpharma-org/drugbank-mcp-server resolves (not archived, pushed 2026-05-07), no OSV advisories, but it is an unofficial community wrapper (not affiliated with DrugBank), the GitHub license classifier is null, and it requires user-supplied license-gated DrugBank XML data. Verification works. Recheck if the repo adds a LICENSE or DrugBank’s terms change.
molecule-mcp, openmm-mcp, rdkit-agent, labmate-mcp, scitex, neuroflow (standalone) — security caution · 2026-07-20 — provenance matches each supplier and no OSV advisories, but each has a maintenance/license risk signal: molecule-mcp (ChatMol/molecule-mcp MIT but stale, last push 2025-04-20), openmm-mcp (PhelanShao/openmm-mcp-server GitHub license NOASSERTION vs page GPLv3 claim, stale 2025-05-31), rdkit-agent (npm rdkit-agent 0.1.1 MIT, Alpha single-maintainer 9 stars, GitHub license classifier null), labmate-mcp (single-maintainer 2 stars, optional free 3rd-party API keys via --setup), scitex (AGPL-3.0-only copyleft, single-maintainer), neuroflow (early Beta v0.2.x single-maintainer 6 stars, bundles its own MCP). All verification works. Recheck maintenance/license next cycle.
tooluniverse-binder-discovery, tooluniverse-drug-drug-interaction (ToolUniverse) — security caution · 2026-07-20 — provenance/Apache-2.0 clear and skill dirs confirmed, but each surfaces an external-service credential dependency: binder-discovery routes docking/generation through external NVIDIA NIM endpoints needing a user NVIDIA_API_KEY; drug-drug-interaction ships a .env.template requesting user API keys for external DBs (BioGRID/DisGeNET/OMIM/USPTO/NVIDIA/BRENDA). Verification works. The other 17 ToolUniverse pages (16 skills + the MCP server) are read-only over public APIs → cleared. Informational caution only; not a provenance mismatch.
covasyn — verification degraded + security caution · 2026-07-20 — hosted endpoint mcp.covasyn.com/mcp responds (406, API-key gated) but the page’s documented npm @covasyn/mcp-client stdio proxy is a confirmed 404 on the npm registry (registry.npmjs.org/-/v1/search?text=covasyn returns 0 packages). Could not fix to a working non-hosted path — the real connection method is behind covasyn.com account login. Examples repo is under a personal account (oliverkraft93-ops), commercial service, API-key credential. Curator/next run: confirm the current stdio client package name (or drop the two stdio-proxy blocks and keep only the hosted HTTP form).
allenbrain — verification degraded + security caution · 2026-07-20 — repo transferred from maflot to the MCPmed org (api.github.com/repos/maflot/allenbrain-mcp reports full_name: MCPmed/allenbrain-mcp). Fixed the git clone command and Sources link to MCPmed/allenbrain-mcp this run. Still no LICENSE, Alpha, single-maintainer. Curator: the page supplier/Supplier-link still say maflot — reconcile to MCPmed if appropriate.
neurosift — security caution · 2026-07-20 — anchor magland/neurosift-mcps resolves and the clone/build install path is current, but the repo has no LICENSE (page already notes), single-maintainer, last push 2025-11-03. Verification works; recheck maintenance + license next cycle.
boltz, biomni (standalone) — security caution · 2026-07-20 — provenance/license clear (boltz-bio MIT + official marketplace; snap-stanford Apache-2.0) but each has a risk signal: boltz ships a BOLTZ_API_KEY to the hosted paid Boltz API (external-service dependency), and biomni’s A1 agent executes LLM-generated code with full system privileges (sandbox it). Not a provenance mismatch — informational caution only.
consensus, adisinsight (standalone connectors) — verification degraded · 2026-07-20 — both resolve in the official anthropics/life-sciences marketplace and provenance matches, but each remote MCP server is subscription/account-gated (Consensus.app account; AdisInsight subscription) so they are functionally unverifiable without credentials. Security cleared. Recheck if a free/eval tier becomes testable.
what-if-oracle (K-Dense) — security caution · 2026-07-20 — provenance/anchor clear but the page states CC BY-NC-SA 4.0 (non-commercial, share-alike) while the collection root is MIT; the non-commercial clause limits reuse. Recheck the skill’s own license/SKILL.md if K-Dense relicenses.
Batch 8 caution set (K-Dense) — security caution · 2026-07-20 — external-service/credential or unstated-license dependencies: research-lookup (PARALLEL_API_KEY→api.parallel.ai, OPENROUTER_API_KEY→openrouter.ai), open-notebook (user-configured external AI providers), modal (external Modal cloud), parallel-web (external parallel-cli + license unstated), infographics (external Google Gemini + license unstated), paper-lookup/latex-posters (license unstated on page; no external creds — lift to cleared if a license is confirmed), and the integration skills ginkgo-cloud-lab, omero-integration, opentrons-integration, protocolsio-integration, labarchive-integration, latchbio-integration (license unstated + external cloud/ELN credentials). Collection root is MIT; recheck each skill dir for its own license to lift the unstated-license ones.
pdf, docx, pptx (K-Dense) — security caution · 2026-07-20 — each skills/<slug>/LICENSE.txt is Anthropic PBC proprietary (“© 2025 Anthropic, PBC. All rights reserved”, no redistribution/derivatives) redistributed in the MIT K-Dense collection while the page claims Free/OSS (“Proprietary..txt has complete terms”). Same pattern as xlsx; curator should correct each Pricing line.
xlsx (K-Dense) — security caution · 2026-07-20 — skills/xlsx/LICENSE.txt is Anthropic PBC proprietary (“all rights reserved”, no redistribution/derivatives) yet the MIT K-Dense collection redistributes it and the page claims Free/OSS. Curator should correct the Pricing line.
scientific-schematics (K-Dense) — security caution · 2026-07-20 — provenance/MIT clear, but the skill sends prompts plus a user API key to external Google Gemini (“Nano Banana 2” / Gemini 3.1 Pro) image services; note the external-service/credential dependency.
scientific-problem-selection (Anthropic) — security caution · 2026-07-20 — plugin resolves in anthropics/life-sciences but repo has no top-level LICENSE; lift to cleared if a LICENSE lands.
uniprot (Augmented-Nature) — security caution · 2026-07-20 — LICENSE file on the repo is restrictive non-commercial (“personal, non-commercial use only”, no redistribution/modification) while package.json and the catalog page both claim MIT. Curator should reconcile the Pricing line; recheck if upstream relicenses.
rowan — security caution · 2026-07-20 — k-yenko/rowan-mcp repo publishes no LICENSE (page already notes this) and the tool ships a user ROWAN_API_KEY to the external paid Rowan cloud; recheck if the MCP repo adds a LICENSE.
phylogenetics, gtars, pyhealth — security caution · 2026-07-20 — K-Dense provenance clears but each skill’s own upstream library license is unstated on the page; lift to cleared if a license is confirmed.
generate-image, exa-search (K-Dense) — security caution · 2026-07-20 — MIT and provenance clear, but each ships a user API key to an external service (generate-image → FLUX/Nano Banana image services, same pattern as scientific-schematics; exa-search → the Exa web API); note the external-service/credential dependency.
primekg, optimize-for-gpu, hugging-science, dnanexus-integration (K-Dense) — security caution · 2026-07-20 — provenance clears but each page states the skill/data license as unstated; hugging-science can also call the external HF Inference API and dnanexus-integration uses DNAnexus cloud creds via dxpy. Lift to cleared if a license is confirmed.
kegg-pathway-analysis (SciAgent) — security caution · 2026-07-20 — provenance matches jaechang-hits and the skill code is CC BY 4.0, but the underlying KEGG data needs a paid commercial license for non-academic use. Verification works. Data-use restriction only — the skill code itself is unrestricted. Same pattern as kegg-database.
cosmic-database, kegg-database, ddinter-database (SciAgent) — security caution · 2026-07-20 — skill code is CC BY 4.0 and provenance clears, but each page’s underlying data carries a restrictive license: COSMIC data is CC-BY-NC-SA-4.0 (non-commercial, registration required); KEGG data needs a paid commercial license for non-academic use; DDInter data is CC BY-NC 4.0 (non-commercial per the NAR 2022/2025 papers) while the page’s Pricing claims CC-BY-4.0. Data-use restriction only — the skill code itself is unrestricted. Curator should reconcile the ddinter Pricing line. Recheck if the data-use terms change.
SciAgent-Skills CC BY 4.0 correction (supersedes prior NOASSERTION caution) · 2026-07-20 — GitHub’s license classifier reports NOASSERTION for jaechang-hits/SciAgent-Skills, but the committed root LICENSE is verbatim CC BY 4.0 (commercial use + redistribution permitted). SciAgent skills therefore clear on provenance/license by default; only flag a page when its underlying data source has its own restriction (see cosmic/kegg above).
Prior: pymol, foldseek-structural-search degraded but fixed in-page; scmcphub ecosystem on caution for staleness — recheck maintenance next cycle.

Deferred — next-run priority

2026-08-17 (latest run, review-budget batch): worked the injected 15-page adjudication digest (acmg-classification, alignment-trimming, alkyl, blatant-why, can-immune, mcptools, medicare-mcp, msa-statistics, msi-detection, multiple-alignment, neuro-mcp, nwb-mcp-server, pbmcpedia, somatic-signatures, structural-alignment) at the 15-page review budget; 9 further digest pages were over budget this run and were intentionally left untouched — they should lead the next worklist. All 15 worked pages graded works; 13 caution, 2 cleared (mcptools, medicare-mcp — see Flagged). No page-content fixes needed this run (all install/launch commands confirmed accurate against primary sources). Priority finding to carry forward: GPTomics/bioSkills is now archived (see Flagged) — recheck next run whether this persists or reverses, and watch for any further bioSkills pages entering the digest under the new archived status. Next run: resume the digest’s next window.
2026-08-10 (latest run, review-budget batch): worked the injected 15-page adjudication digest (bayesian-trials, cdisc-data-handling, chemgraph, clustering-phenotyping, compartment-analysis, compensation-transformation, covariation-analysis, cytometry-differential-analysis, cytometry-qc, effect-measures, fda-mcp, gating-analysis, geometric-analysis, hashing-demultiplexing, lesion-symptom-mapping-guide) at the 15-page review budget; 17 further flagged pages were over budget this run and were intentionally left untouched — they should lead the next worklist. All 15 worked pages graded works; 14 cleared, 1 caution (cdisc-data-handling — see Flagged). No page-content fixes needed this run (all evidence — GPTomics MIT root, chemgraph Apache-2.0, fda-mcp MIT, lesion-symptom-mapping-guide repo-renamed — confirmed pages already accurate). Next run: resume the digest’s next window; no open follow-ups from this batch beyond normal recheck cadence.
2026-08-06 (latest run, review-budget batch): worked the injected 15-page adjudication digest (omophub-mcp, optogenetics-protocol-designer, perturb-seq, pyomop, rosetta-mcp-server, scatac-analysis, signal-detection-analysis, strain-tracking, structure-preparation, structure-validation, tooluniverse-admet-prediction, tooluniverse-cell-line-profiling, tooluniverse-chemical-sourcing, tooluniverse-dose-response, trial-reporting) at the 15-page review budget; one further flagged page from the digest was over budget this run and was intentionally left untouched — it should lead the next worklist. All 15 worked pages graded works or degraded; one content fix applied (pyomop launch command — see Flagged). Priority follow-ups: (1) pyomop — confirm pyomop --mcp-server boots cleanly in a fresh smoke test and flip degraded→works; (2) omophub-mcp — recheck if a free/no-signup tier becomes testable; (3) general cadence — no other open follow-ups from this batch.
2026-08-03 (latest run, review-budget batch): worked the injected 15-page adjudication list (adaptive-designs, binding-site-detection, calcium-imaging-analysis-guide, cnv-inference, deeplabcut, differential-abundance, doublet-detection, drift-diffusion-model, functional-profiling, immunogenicity-scoring, interface-analysis, lineage-tracing, materials-project-mcp, metaphlan-profiling, missing-data-sensitivity) out of a 31-page liveness digest; 16 further flagged pages were over this run’s review budget and stay due, leading the next worklist (not stamped this run). All 15 worked pages graded works; 14 cleared, 1 caution (immunogenicity-scoring — see Flagged). No page-content fixes were needed this run (all three flags — repo-renamed ×3, smoke-install-error ×1 — resolved as non-issues on inspection, see Recently verified for detail). Next run: resume the digest’s next window; no open follow-ups from this batch beyond normal recheck cadence.
2026-07-29 (prior run): worklist advanced to a 25-page window (umap-learn → aind-data) — 24 clean rechecks + 1 launch-command fix (aind-data → HTTP transport, degraded; see Flagged). Stamps refreshed 2026-07-20→2026-07-29. Next run: keep working the injected worklist top-to-bottom; carry the follow-ups below. Priority follow-up from this run: recheck aind-data’s HTTP endpoint (metadata-portal.allenneuraldynamics.org/mcp/) — flip degraded→works once the HTTP transport proves stable across a run. Cadence-watch: the K-Dense/SciAgent/NeuroClaw anchors are all fresh (pushed within the last week); re-fetch each collection LICENSE only if a future push changes the root license. Kept-caution set on this batch unchanged (xlsx Anthropic-proprietary, what-if-oracle CC BY-NC-SA, adaptyv wet-lab API key, uniprot restrictive-LICENSE-vs-MIT-claim).
2026-07-27 (prior run): worklist advanced again to a 4-page window (latchbio-integration → liana-mcp) — all four were clean rechecks (zero drift), stamps refreshed 2026-07-20→2026-07-27. latchbio-integration + latex-posters stay works/caution (K-Dense MIT root, per-skill license unstated on page; latchbio also ships external LatchBio cloud creds); lggnn works/cleared (NeuroClaw MIT); liana-mcp works/caution (scmcphub no-LICENSE, unmaintained since 2025-06, liana-mcp run launch confirmed in README). Next run: keep working the injected worklist top-to-bottom and carry the follow-ups below.
2026-07-27 (prior run): worklist advanced again to a 4-page window (kegg-database → kmeans) — all four were clean rechecks (zero drift), stamps refreshed 2026-07-20→2026-07-27. Worklist header reports 459 total · 0 unstamped — full verification: coverage, every future run is a rolling recheck. kegg-database + kegg-pathway-analysis stay works/caution (KEGG data paid commercial license for non-academic use; skill code CC BY 4.0 clean); ketcher works/cleared (epam/ketcher Apache-2.0 pushed 2026-07-27, npm ketcher-react 3.17.1); kmeans works/cleared (NeuroClaw MIT). Next run: keep working the injected worklist top-to-bottom and carry the follow-ups below.
2026-07-27 (prior run): worklist advanced to a 4-page window (inductive-bio → intact) — all four were clean rechecks (zero drift), stamps refreshed 2026-07-20→2026-07-27. inductive-bio stays degraded/cleared (no public endpoint; not in Claude Science doc, confirmed only via PR Newswire); the other three unchanged.
2026-07-27 (earlier run): worklist advanced to a 4-page window (hypothesis-crucible → ica) — all four were clean rechecks (zero drift), stamps refreshed 2026-07-20→2026-07-27. Next run: keep working the injected worklist top-to-bottom and carry these two follow-ups from the earliest 2026-07-27 pass —
2026-07-27 (earlier run): worklist advanced (admetlab-mcp → npi-registry) — the 12 unstamped pages were all stamped, bringing the catalog to full coverage. Next run: keep working the injected worklist top-to-bottom and prioritize these two fresh follow-ups — (1) novomcp: fetch the new NovoMCP/novomcp README, confirm the localhost:8018/mcp self-host launch command, reconcile LICENSE (NOASSERTION vs Apache-2.0 site claim) — likely flips degraded/unknown → works/caution; (2) drug-pipeline-mcp: recheck PyPI for drug-pipeline-mcp (404 on all 3 endpoints this run) — if it appears, flip degraded → works; if still absent, hand the pip/uvx-vs-git-source discrepancy to the curator. Also, cdxml-toolkit and chimerax-mcp are good smoke-queue candidates ONLY if the sandbox can satisfy their GUI/ChemDraw/ChimeraX prerequisites — otherwise leave them as static-confirmed works.
RESOLVED 2026-07-22: the selector loop is broken — this run served a NEW window (ketcher → medical-terminologies-mcp) instead of the stuck 10x-genomics-cloud→autodock 25 that re-served five times. The same-date rotation the maintainer added is advancing the pointer, so coverage is now progressing across the ~447-page catalog on a rolling basis. Next run: keep working whatever fresh worklist is injected top-to-bottom; the stuck batch is behind us. All 25 pages this window were clean rechecks (zero drift). Note for cadence: K-Dense-AI/scientific-agent-skills had a fresh push (2026-07-20→2026-07-21) — provenance/MIT unchanged, but re-fetch its LICENSE/skill dirs if a future push changes the collection license.
DONE 2026-07-20 (worklist maintenance batch #5, same 10x-genomics-cloud → autodock-vina-docking list served a FIFTH time): all 25 rechecked again against fresh source fetches — zero drift, no fixes. Fifth consecutive identical batch; the selector loop is unchanged and still self-perpetuating. Maintainer action still needed (verifier cannot self-correct): scripts/select_verify_targets.py uses a stable verified_on-oldest tie-break, and because the whole catalog is uniformly dated 2026-07-20 the same 25 slugs sort first every run, so the pointer never advances and the other ~420 pages are never rechecked. Fix options unchanged from batch #3/#4: add a secondary tie-break to the selector (e.g. round-robin hash of slug against run date/commit) OR have the workflow advance a rotating cursor so successive runs serve different windows. Until the selector changes, every maintenance run keeps re-verifying only these 25 with a green “complete” count while coverage silently stalls — exactly the blind-spot class VERIFIER_AGENT.md warns about. This run’s anchor re-fetches (all unchanged, backing existing grades): NeuroClaw MIT/2026-07-14/75★, K-Dense MIT/2026-07-20/31.3k★, DeepMind science-skills Apache-2.0/2026-07-07/2469★, GPTomics bioSkills MIT/2026-07-18/1042★, DeepMind alphafold Apache-2.0/2026-04-22/14.7k★, Augmented-Nature AlphaFold+BioStudies NOASSERTION/2025-12-21, SciAgent-Skills NOASSERTION-classifier-CC-BY-4.0-root/2026-06-15/278★, PyPI arxiv-mcp-server 0.5.1 + aind-data-mcp 0.4.5, MCPmed/allenbrain-mcp no-LICENSE/2026-04-01/3★, life-sciences marketplace still lists adisinsight+10x-genomics.
DONE 2026-07-20 (worklist maintenance batch #4, same 10x-genomics-cloud → autodock-vina-docking list served a FOURTH time): all 25 rechecked again against fresh source fetches — zero drift, no fixes. This is now the fourth consecutive identical batch. The selector loop is fully confirmed: select_verify_targets.py uses a stable verified_on-oldest tie-break, and because the whole catalog is uniformly dated 2026-07-20 the same 25 slugs sort first every run, so the pointer never advances and the other ~420 pages are never rechecked. Maintainer action needed — this run cannot self-correct the loop (the verifier must not bump verified_on beyond the run date, and all pages already carry today’s date): add a secondary tie-break to the selector (e.g. round-robin hash of slug against the run date/commit) OR have the workflow advance a rotating cursor so successive runs serve different windows. Until the selector changes, every maintenance run will keep re-verifying only these 25 with a green “complete” count while coverage silently stalls — exactly the blind-spot class VERIFIER_AGENT.md warns about.
DONE 2026-07-20 (worklist maintenance batch #3, same 10x-genomics-cloud → autodock-vina-docking list served a THIRD time): all 25 rechecked again against fresh source fetches — zero drift, no fixes. The selector keeps re-serving the identical 25 because the whole catalog is uniformly dated 2026-07-20, so verified_on-oldest ties resolve to the same slug ordering every run. This is now a confirmed selector loop: the select_verify_targets.py tie-break is stable, so a uniformly-dated catalog re-serves the same batch indefinitely. Recommendation for the maintainer: add a secondary tie-break (e.g. round-robin by hashing slug against run date) or advance verified_on on rechecked pages so the pointer moves — otherwise the other ~420 pages never get rechecked. Until then, treat any page as a fair recheck target; the three rechecks of these 25 have all confirmed unchanged.
DONE 2026-07-20 (worklist maintenance batch #2, same 10x-genomics-cloud → autodock-vina-docking list): all 25 rechecked again against fresh source fetches; 24 unchanged, 1 micro-fixed (arxiv security_note version 0.5.0→0.5.1, grade unchanged). The selector re-served the identical batch because these are the 25 oldest verified_on (all 2026-07-20) — next run should still resume from the top of a freshly-computed worklist; if it keeps re-serving the same 25, the whole catalog is now uniformly dated 2026-07-20 and any page is a fair recheck target.
DONE 2026-07-20 (worklist maintenance batch #1, 10x-genomics-cloud → autodock-vina-docking): all 25 rechecked against fresh source fetches; 24 unchanged, 1 fixed (autodock-vina-docking caution→cleared — SciAgent root LICENSE is CC BY 4.0, superseding the stale NOASSERTION caution). Next run: resume from the top of a freshly-computed worklist (verified_on-oldest first). Cadence-watch items on this batch — alphagenome/adisinsight/10x-genomics-cloud stay degraded (signup/subscription/paid gated; retest for a free tier); the Augmented-Nature pair (alphafold, arrayexpress) and allenbrain stay caution (restrictive/absent LICENSE); recheck K-Dense astropy/aeon/arboreto/anndata BSD-3-wrapper claims only if K-Dense relicenses. Lesson reconfirmed: any jaechang-hits/SciAgent-Skills page whose security_note cites GitHub NOASSERTION is stale — the committed root LICENSE is CC BY 4.0, so those clear by default unless the page overstates its own data-source license.
biomcp — recheck next clean run to flip degraded→works: the launch command was corrected to biomcp serve this run (auto-fix ⇒ degraded per rubric). Nothing else off; verify the four occurrences still read serve and stamp works. New guardrail now live — the launch-command static check catches this whole class (a package resolves but its documented invocation is dead/renamed), so it applies to every future MCP-server/CLI stamp, not just biomcp.
DONE 2026-07-20 (launch-command sweep): every catalog entry with a trailing launch subcommand verb was validated against upstream; only biomcp was wrong (fixed). scmcphub run, bci-mcp serve, rdkit-agent mcp, chatspatial server, scitex mcp start all confirmed correct. Remaining lower-risk launch shapes not individually re-fetched this pass (entrypoint/module paths — python -m <mod>.server for blast/gwas-mcp/spikelab, uv run ... start for fhir-momentum, mcp run ./server.py for chemlint): re-confirm these against each repo’s README on the normal maintenance cadence.
DONE 2026-07-20 (worklist batch): the 7 unstamped worklist pages are all stamped — mygene (works/cleared); alphafold, gene-ontology, brian2, clinical-trial-protocol, cms-coverage (works/caution); biorender (degraded/cleared). The 8 already-stamped worklist items were rechecked against fresh source fetches (NeuroClaw MIT/maintained; K-Dense MIT/maintained + smoke-installable) and left at their existing 2026-07-20 grades. Curator handoff: alphafold + gene-ontology Pricing rows were falsely claiming MIT/OSS for the Augmented-Nature wrapper code, which is actually a restrictive personal non-commercial LICENSE (GitHub NOASSERTION) — verifier fixed the Pricing lines this run; curator owns any further wording. Same license pattern now confirmed across the Augmented-Nature MCP-server family (uniprot, human-protein-atlas, alphafold, gene-ontology, arrayexpress/BioStudies, OpenTargets fallback) — treat any other Augmented-Nature/*-MCP-Server page’s MIT/OSS claim as suspect and fetch its raw LICENSE before stamping. Next run: resume from the top of a freshly-computed worklist (bootstrap should now be complete-or-near — confirm coverage) and keep rechecking the degraded/broken items below on cadence.
DONE 2026-07-20 (maintenance recheck pass 3): re-fetched all five open priority items + two staleness-cadence caution items, none regraded — openneuro still broken (repo + endpoint 404 for the 4th consecutive run; refreshed note/flag/table row to “fourth consecutive run” — curator: strong signal to remove/replace), glygen still 503, open-targets source repo opentargets/platform-mcp still Apache-2.0/current + official endpoint still GET-untestable, covasyn npm @covasyn/mcp-client still 0 results, morning still absent from anthropics/skills/skills/. Staleness rechecks unchanged (notes already accurate): scanpy (scmcphub/scanpy-mcp BSD-3 pushed 2025-06-27), blast (bio-mcp/bio-mcp-blast license null pushed 2025-06-29). Coverage re-confirmed 440/440. Next cycle: keep rechecking these on cadence (grade glygen broken if 503 persists AND self-host repo regresses; retest open-targets official endpoint with a real MCP client; watch novomcp for a public client/free tier) and lift the no-LICENSE anthropics/healthcare skills to cleared if a LICENSE lands.
DONE 2026-07-20 (maintenance recheck): openneuro regraded broken + flagged (repo + hosted endpoint 404 for the 2nd consecutive run — hand off to curator for removal/replacement). Confirmed-unchanged rechecks (leave dated 2026-07-20, revisit on cadence): glygen (hosted endpoint still 503 — retest next cycle; consider degraded→broken if the hosted endpoint stays down AND the self-host repo regresses), open-targets (official endpoint initialize still untestable via GET WebFetch — retest with a client; source repo redirect opentargets/open-targets-platform-mcp→opentargets/platform-mcp still resolves Apache-2.0), covasyn (hosted 406 API-key-gated, npm @covasyn/mcp-client still absent — curator to reconcile the stdio client package name), novomcp (recheck for a public client/free tier). Still open on the LICENSE-lands cadence: the no-LICENSE anthropics/healthcare skills (fraud-detection/procedure-coding/clinical-note-extract/prior-auth-review/icd-10-codes/scientific-problem-selection) and morning (locate the correct plugin/bundle). Smoke-note refresh done for rdkit-skill+scikit-bio (git ENOENT resolved; both K-Dense slugs now pass via npx skills add).
DONE 2026-07-20 (pass 30): the final 16 unstamped pages are all stamped (npi-registry, pubmed, fhir-wso2, proto-okn, aind-data, mhc-binding-prediction, drug-repurposing, hypothesis-crucible → works/cleared; fraud-detection, procedure-coding, clinical-note-extract, openfda, cbioportal, arrayexpress → works/caution; bio-research, 10x-genomics-cloud → degraded — see Flagged). A full-catalog sweep confirms 440 of 440 tool pages now carry ^verification: — the bootstrap enumeration is complete. Next-run priority shifts to RECHECKING the degraded/caution pages on their own cadence rather than fresh enumeration: openneuro (repo + endpoint 404 — if still 404 grade broken + set flagged:), morning (locate correct plugin/bundle), open-targets (retest official MCP initialize), glygen/novomcp/covasyn (retest hosted endpoints), and the no-LICENSE anthropics/healthcare skills (fraud-detection/procedure-coding/clinical-note-extract/prior-auth-review/icd-10-codes/scientific-problem-selection) — lift to cleared if a LICENSE lands. Enumeration reminder: Grep count gives false “Found 0” for line-1 matches; use files_with_matches count (440) vs the LS file total to detect any newly-added unstamped pages.
DONE 2026-07-20: the 12 previously-deferred tail pages are all stamped (tcr-epitope-binding, single-cell-rna-qc, pdbe, medical-terminologies-mcp, indication-dossier → works/cleared; pynibs, ontology-lookup-service, prior-auth-review, icd-10-codes → works/caution; open-targets, openneuro, morning → degraded — see Flagged). Next-run priority: RECHECK the three degraded pages — openneuro (repo + hosted endpoint both 404 this run; if still 404 next run grade broken + set flagged:), morning (locate the correct plugin/bundle or reconcile the Claude Science claim), open-targets (retest the official MCP initialize endpoint). Then continue enumerating any still-unstamped catalog pages: Grep catalog/tools/*.md for files lacking ^verification: (Bash multi-op is gated; the Task/subagent tool 404’d on the sonnet model previously — use Grep pattern compares to enumerate). Confirm each page’s supplier: + anchor before stamping.
Standalone MCP / plugin / connector batch — DONE 2026-07-20 (see Recently verified): 15 unstamped non-K-Dense/non-SciAgent pages stamped — rdkit-mcp, spikelab, seqera, healthlake-mcp, scholar-gateway, synapse (works/cleared); molecule-mcp, openmm-mcp, rdkit-agent, labmate-mcp, scitex, drugbank, neuroflow (works/caution); novomcp, cortellis (degraded). Remaining standalone unstamped candidates for next run: blast (bio-mcp org, recheck LICENSE), scanpy/decoupler-mcp/liana-mcp/cellrank-mcp (scmcphub, smoke-queued), plus any other non-K-Dense/non-SciAgent MCP servers surfaced by an LS sweep of catalog/tools/.
NeuroClaw supplier — DONE 2026-07-20 (all pages stamped): this pass stamped the final 6 (qsiprep-tool, run_models, spacenet, wmh-segmentation, nii2dcm, nilearn-tool) on top of the earlier 25 + the *-skill.md dataset pages. nii2dcm note reflects MIT skill code wrapping BSD-3-Clause upstream. No NeuroClaw pages remain unstamped.
Anthropic tool_type: Claude.ai Connector pages — DONE 2026-07-20 (all 21 stamped this pass; see Recently verified). The 16 Anthropic-hosted featured-connector data sources + ketcher = works/cleared; medidata = works/cleared (marketplace + published endpoint); owkin/revvity-signals/inductive-bio = degraded/cleared (provenance confirmed, no publicly resolvable endpoint). No connector pages remain unstamped.
This run stamped SciAgent non-database skill batch 4 (15 supplier: SciAgent skills: mdanalysis-trajectory, sar-analysis, smina-molecular-docking, multiqc-qc-reports, plannotate-plasmid-annotation, sgrna-design-guide, libsbml-network-modeling, plotly-interactive-plots, napari-image-viewer, opencv-bioimage-analysis, cellpose-cell-segmentation, scikit-image-processing, celltypist-cell-annotation, harmony-batch-correction, simpleitk-image-registration — all works/cleared) plus the 4 GPTomics bioSkills pages (neoantigen-prediction, epitope-prediction, scirpy-analysis, mhc-class-ii-prediction — anchor GPTomics/bioSkills MIT, all works/cleared). SciAgent anchor re-confirmed jaechang-hits/SciAgent-Skills (CC BY 4.0 root LICENSE, pushed 2026-06-15, updated 2026-07-20; new parent dir data-visualization/ confirmed this run). Parent dirs STILL hold many more unstamped SciAgent skills — confirm each catalog page is supplier: SciAgent AND its parent contents dir before stamping:
- structural-biology-drug-discovery/ remaining: pubchem-compound-search, pdb-database (some may be -database or already stamped — check supplier: and existing stamp).
- genomics-bioinformatics/qc/ remaining: busco-status-interpretation.
- lab-automation/ remaining: opentrons-protocol-api, pylabrobot (check supplier — pylabrobot may be K-Dense).
- medical-imaging/ remaining: histolab-wsi-processing, pydicom-medical-imaging (some are K-Dense pages — check supplier).
- proteomics-protein-engineering/ remaining: adaptyv-bio, esm-protein-language-model, matchms-spectral-matching, pyopenms-mass-spectrometry (check supplier).
- systems-biology-multiomics/ remaining: cobrapy-metabolic-modeling, lamindb-data-management (check supplier — lamindb/cobrapy have K-Dense pages).
- scientific-computing/ (aeon, astropy-astronomy, dask-parallel-computing, exploratory-data-analysis, geopandas-geospatial, hypogenic-hypothesis-generation, matlab-scientific-computing, networkx-graph-analysis, neurokit2, neuropixels-analysis, nextflow-workflow-engine, polars-dataframes, pyhealth, pymatgen, pymoo, scikit-learn-machine-learning, shap-model-explainability, simpy-discrete-event-simulation, sympy-symbolic-math, torch-geometric-graph-neural-networks, transformers-bio-nlp, umap-learn, vaex-dataframes, zarr-python — many have separate non-SciAgent catalog pages, check supplier: before stamping), cell-biology/ remaining (flowio-flow-cytometry), genomics-bioinformatics/single-cell/ remaining (anndata-data-structure, cellxgene-census, scanpy-scrna-seq, scvi-tools-single-cell), data-visualization/ remaining (any non-plotly siblings — confirm via contents API). Watch for per-skill data/tool license restrictions (only flag if the page overstates its license, e.g. kegg).
Prior run stamped the entire ToolUniverse family (19 pages: 1 MCP server + 18 skills) against mims-harvard/ToolUniverse Apache-2.0 + PyPI tooluniverse 1.0.22 — all works, 17 cleared and 2 caution (binder-discovery, drug-drug-interaction; see Flagged). Tooling reminders that held again: OSV api.osv.dev/v1/query is POST-only (WebFetch GET → 405) so use GitHub repos/<org>/<repo>/security-advisories (GET); the Read cache invalidates across files AND a MultiEdit right after a fresh Read can trip the permission gate — the reliable pattern is Read-then-two-single-Edits per file, stamped strictly sequentially.
Prior run (pass 15) stamped a second non-K-Dense standalone batch: evo2, scgpt, proteinmpnn, solublempnn, openfold3, chatspatial, biocontextai (works/cleared), neurosift (works/caution), allenbrain + covasyn (degraded/caution; see Flagged).
Good next standalone batches (still many unstamped non-K-Dense pages): the remaining supplier: SciAgent skill pages that are NOT -database (e.g. deseq2-differential-expression, star-rna-seq-aligner, gatk-variant-calling, snakemake-workflow-engine, the many *-analysis/aligner/annotation skills) — same anchor jaechang-hits/SciAgent-Skills (CC BY 4.0 root LICENSE) but confirm each skill’s parent dir via contents API; the GPTomics bioSkills pages are DONE (all 4 stamped 2026-07-20); Anthropic anthropics/life-sciences skill/connector pages (metabolights, complex-portal, rfam, intact, bindingdb, clinical-trial-protocol, finngen, etc. — verify each in the marketplace’s .claude-plugin/marketplace.json + <name>/.claude-plugin/plugin.json); and standalone repo/PyPI-anchored servers (biomcp and tooluniverse family done; blast, scanpy, drugbank, openneuro).
Next bootstrap batch: work the remaining unstamped supplier: K-Dense catalog pages, confirming each skills/<slug> dir with a direct contents/skills/<slug> fetch (the summarized listing is unreliable in both directions). Then continue with remaining non-K-Dense unstamped suppliers (standalone MCP servers / skills) — a large share of the unstamped pages are outside the K-Dense anchor. Editing lesson (holds every run): stamp each file strictly sequentially (read-then-two-edits per file); parallel edits across files trip the read-cache/permission gate.
Note (superseded 2026-07-20): a prior note claimed scipy, seurat, squidpy, spatialdata, survival-analysis, systems-biology, tensorflow, torch etc. were absent from the K-Dense skills/ tree. This run’s clean tail fetch shows several of those names (scipy, seurat, squidpy, spatialdata, survival-analysis, systems-biology, tensorflow) DO appear in the listing. Given the summarized listing is unreliable in both directions, do NOT treat any slug as present or absent from the list alone — confirm with a direct contents/skills/<slug> fetch before stamping (or declining) each catalog page.
Curator handoff: xlsx, pdf, docx, pptx page Pricing lines falsely claim Free/OSS (each skill’s upstream LICENSE.txt is Anthropic PBC proprietary) — verifier stamped security caution; curator owns the Pricing corrections.
scmcphub ecosystem (scanpy, cellrank-mcp, decoupler-mcp, liana-mcp) — recheck for a maintenance bump or archival; currently ~13 months stale.
gromacs-mcp — recheck upstream for a published LICENSE file; graded caution until then.
blast (bio-mcp org) — recheck for a published LICENSE and a maintenance bump; caution until then.
SciAgent -database remainder (bootstrap) — RESOLVED this run (batch 2): biorxiv, interpro, hmdb, gtopdb, emdb, brenda, unichem, uspto, zinc, reactome, pride, openalex, opentargets, metabolomics-workbench, dailymed, fda, ddinter all stamped, plus string-database-ppi. Key lesson: these skills do NOT all live under skills/genomics-bioinformatics/databases/ — they are scattered across structural-biology-drug-discovery/, proteomics-protein-engineering/, systems-biology-multiomics/, scientific-writing/, scientific-computing/ (the path is in each page’s Notes line “The skill directory upstream is skills/<...>”). Confirm each by fetching that parent contents endpoint. Still to do: the Google DeepMind unibind/gtex pages (different anchor google-deepmind/science-skills; verify separately). NOTE: ensembl-database.md has no catalog page (present in repo listing only) — do not chase it.
SUPERSEDED 2026-07-20: the old note that SciAgent-Skills entries (bcftools-variant-manipulation, autodock-vina-docking, …) are caution due to GitHub NOASSERTION is wrong — the committed root LICENSE is CC BY 4.0. Those Augmented-Nature/SciAgent skill pages clear on provenance/license by default; recheck each only for its own data-source or per-skill license.
chemcp — recheck for a LICENSE file committed to the repo to reconcile the ISC npm metadata.

Smoke-test queue

The scripts/select_smoke_targets.py selector is authoritative for what the quarantined smoke job may install/boot (open, no-auth, no-cost Skills/MCP servers only). List slugs here to prioritize the next run’s smoke batch; leave empty to let the selector pick by age.

Note (2026-07-27): the latest batch is 12/12 pass — openfda-mcp-server (npx), bci-mcp (npx), biomcp (uv tool), anndata/arboreto (pip), and the K-Dense CLI batch (npx skills add K-Dense-AI/scientific-agent-skills, several aeon/astropy/bids slugs) all install/boot cleanly. The lone non-pass was cdxml-toolkit boot_error, and that is a false negative — the resolver backtracked to v0.5.1 (no cdxml-mcp entry point); the current v0.5.17 metadata does carry the entry point, so the page stays works. Node in-sandbox is v20.19.2 (skills@1.5.20 wants >=22.20.0) — harmless EBADENGINE warning, install still succeeds. Next smoke priorities:

alkyl (Kdevos12) — no smoke result yet; static-only works this run, real install would upgrade the evidence.
blatant-why (001TMF) — no smoke result yet; check whether a no-key code path exists before queuing (ships optional external API keys).
medicare-mcp (openpharma-org) — no smoke result yet; no published npm package, would need a git clone && npm install && npm run build smoke path rather than a plain install.
nwb-mcp-server (Ben Hardcastle) — no smoke result yet; uvx nwb-mcp-server --root_dir data --glob_pattern "*.nwb" needs a sample .nwb fixture to boot meaningfully.
pbmcpedia (MCPmed) — no smoke result yet; git clone && npm install && npm start -- --transport stdio, no API key required.
scanpy (pip install scanpy-mcp)
decoupler-mcp (pip install decoupler-mcp)
liana-mcp (pip install liana-mcp)
cellrank-mcp (pip install cellrank-mcp)
protein-mcp-server (npx -y @cyanheads/protein-mcp-server@latest) — stamped works by static bin resolution this run; a boot smoke would upgrade the evidence to a real verdict.
Aging K-Dense slugs — CLI batch confirmed installable in-sandbox; keep rotating fresh slugs to capture boot verdicts.
From the 2026-08-03 batch, smoke_eligible: true / smoke_status: null in liveness.json (bioSkills Python skills — no external creds): binding-site-detection, cnv-inference, doublet-detection, functional-profiling, immunogenicity-scoring, interface-analysis, metaphlan-profiling — good candidates for the next smoke rotation.
lineage-tracing — recheck the smoke job’s target string next run; the current smoke installs the stale/deprecated pip install cassiopeia-lineage (PyPI 1.0.4, install_error on Py3.12) instead of the page’s actual recommended pip install git+https://github.com/YosefLab/Cassiopeia@master — false negative, do not regrade from this alone; consider updating the selector’s target to the git install.
NOT smoke candidates: cdxml-toolkit (Windows + ChemDraw), chimerax-mcp (GUI + ChimeraX) — queue only if the sandbox can satisfy those GUI/native prerequisites.
From the 2026-08-10 batch, smoke_eligible: true / smoke_status: null in liveness.json (bioSkills R/Python skills — no external creds): clustering-phenotyping, compartment-analysis, compensation-transformation, covariation-analysis, cytometry-differential-analysis, cytometry-qc, effect-measures, gating-analysis, geometric-analysis, hashing-demultiplexing — good candidates for the next smoke rotation (note several need R/Bioconductor installs, not just pip).