NWB MCP Server

MCP server that exposes Neurodata Without Borders (NWB) files — on disk, on S3, or in a DANDI dandiset — as a virtual SQL database Claude can explore and query read-only.

   
Type MCP server
Supplier Ben Hardcastle
Availability Beta (PyPI 0.1.9, released 2026-06-07; repo last pushed 2026-06-07)
Pricing Free / OSS — MIT
Capabilities Read-only — the server never writes to NWB files
Verified works · 2026-08-17
Security caution · 2026-08-17 — MIT and provenance confirmed; uv OSV advisories all fixed-before pinned version, non-issue; single-maintainer project pinning a pre-release dependency

How to install

Requires Python 3.10+ and uv (pip install uv). No clone is needed — uvx fetches the published package.

  • Verify it starts (one-shot; Ctrl-C once it boots — Claude Code and Claude Desktop launch the process themselves over stdio): 
    uvx nwb-mcp-server --root_dir data --glob_pattern "*.nwb"

    (replace data with the directory holding your NWB files — e.g. /Users/you/data/nwb, or $(pwd)/data.)

  • Claude Code — direct MCP add (stdio): 
    claude mcp add --transport stdio nwb -- uvx nwb-mcp-server --root_dir /absolute/path/to/nwb-files --glob_pattern "*.nwb"

    (replace /absolute/path/to/nwb-files with the absolute path of the directory containing your .nwb files.)

  • Claude Code — point it at a DANDI dandiset instead of local files: 
    claude mcp add --transport stdio nwb-dandi -- uvx nwb-mcp-server --dandiset_id 000363 --anon
  • Claude Desktop — add to claude_desktop_config.json: 
    {
  "mcpServers": {
    "nwb": {
      "command": "uvx",
      "args": [
        "nwb-mcp-server",
        "--root_dir", "/absolute/path/to/nwb-files",
        "--glob_pattern", "*.nwb"
      ]
    }
  }
}

    (uvx must be on the PATH Claude Desktop sees; if it is not, use the absolute path to the uvx binary, e.g. /Users/you/.local/bin/uvx.)

What it does

Builds a lazy, schema-inferred virtual table view over a collection of NWB files using lazynwb, then lets Claude interrogate it with SQL rather than by generating and running analysis scripts. Ten read-only tools:

  • Discovery — get_tables, get_table_schema, get_nwb_paths, preview_table_values
  • Query — execute_query (SQL against the virtual NWB database)
  • Source switching — get_active_source, use_local_source, use_dandiset_source, reset_active_source
  • Escape hatch — nwb_file_search_code_snippet, which returns Python for locating files when SQL is not enough

Configuration flags include --root_dir, --glob_pattern, --dandiset_id, --dandiset_version, --dandiset_path_filter, --tables, --infer_schema_length, --anon (anonymous S3), --unattended, and --max_result_rows (default 50).

Primary use cases: rapid exploration of an unfamiliar NWB dataset, cross-file unit/trial table queries, generating summary reports over a dandiset.

Notes

  • Read-only by design. The server has no write path into NWB files, which is the main reason to prefer it over letting Claude run pynwb code directly.
  • Schema inference reads one file by default (--infer_schema_length 1). If files in a collection have heterogeneous tables or columns, raise this or the schema will be incomplete.
  • Result rows are capped at --max_result_rows (50) and the README documents a table_element_limit of 500 elements (columns × rows) — large result sets are truncated rather than streamed, so aggregate in SQL rather than pulling raw traces.
  • SQL, not signal processing. This surfaces the tabular side of NWB (units, trials, epochs, electrodes, subject metadata). Continuous acquisition data and spike-waveform analysis are out of its remit — pair it with SpikeLab, SpikeInterface, or Neurosift Tools MCP (which covers DANDI semantic search and dandiset_info).
  • Upstream flag spelling. The README’s JSON example uses underscore flags (--root_dir, --glob_pattern); the snippets above follow it. The upstream docs are written for VS Code Copilot Chat, and state only that “similar agent extensions, such as Cline or Claude Code, should also be able to connect to the server” — the Claude Code and Claude Desktop forms above are adapted from the published command/args pair. Unverified — the maintainer does not publish Claude-specific registration snippets.
  • lazynwb is pinned to a pre-release (lazynwb==1.0.0dev3), so expect API churn.
  • Experimental, undocumented support exists for non-NWB tabular files (CSV, Parquet) via Polars I/O.
  • Small project: 2 GitHub stars as of 2026-08-16. The maintainer works on Allen Institute neuropixels data tooling, which is the workload the server is shaped around.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.