neuro-mcp

A MCP for NeuroAgents that assist clinicians and researchers. It gives an AI agent one interface over the whole workflow: signal processing and source imaging (via MNE-Python), a persistent dataset + EHR store (Postgres + BIDS), and NeuroII web visualization.

It builds on the processing core of eeg-mcp (copied in and rebranded eeg→neuro) and adds the data, EHR, and visualization layers around it.

Architecture

                       neuro-mcp  (FastMCP "neuro-analysis", 54 tools)
 ┌──────────────────────┬────────────────────────┬──────────────────────┐
 │ processing            │ data + EHR store       │ neuroii              │
 │ (MNE + ESI)           │ (Postgres + BIDS)      │ (web visualization)  │
 └──────────────────────┴────────────────────────┴──────────────────────┘
   in-memory                SQLAlchemy +               HTTP client
   session                  BIDS-on-disk               (stub contract)

Actors & workflows

• Clinician — reviews a recording, adds/edits annotations, and amends EHR (records a diagnosis/observation, corrects a value), then signs off.
• Researcher — discovers datasets, imports to BIDS, runs MNE processing + source imaging.
• Agent — orchestrates the above via tool calls.

Clinical-safety model (EHR & annotations)

EHR records and annotations are versioned, never overwritten or hard-deleted:

• Amend = a new audited version. amend_ehr_record / update_annotation insert a new version; the prior one is retained with status amended. So a clinician can modify the EHR — the current view updates while the original and its author are preserved.
• Retract = soft void. void_ehr_record / void_annotation set status entered-in-error; the record stays in the history.
• Every mutation is audited (audit_log: actor, action, before/after).
• Mutating tools take an explicit actor so authorship is on the record. (Auth/RBAC enforcement is planned for v0.2; the fields and trail are in place.)

Each tool returns an outcome field for the operation (created/amended/voided/…) distinct from the record's clinical status, so the two never collide.

Tools (54)

• Processing (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_*) and source imaging / ESI (fetch_template_head … extract_label_timecourses).
• Data/EHR: register_subject, get_subject, add_ehr_record, amend_ehr_record, get_ehr_history, void_ehr_record; import_recording, register_dataset, query_datasets, list_recordings; add_annotation, update_annotation, list_annotations, void_annotation; get_audit_log.
• neuroii: neuroii_push_recording, neuroii_create_viz_session, neuroii_pull_annotations.
• neuroii visualizations (standalone interactive HTML, Plotly): visualize_timeseries (stacked multi-channel EEG with scroll + amplitude buttons), visualize_averaging (ERP butterfly + scalp topomap scrubbed by a time slider), visualize_esi (source-estimate ROI time courses + per-time activation bars).

Install

conda activate eeg-mcp            # or any Python >=3.10 env
pip install -e .                  # core
pip install -e ".[postgres]"      # + PostgreSQL driver (LGPL-3.0)
pip install -e ".[viz3d]"         # + 3D source rendering (PySide6, LGPL-3.0)

Configure (environment variables)

The default (SQLite + a scratch BIDS dir) runs with zero setup; point DATABASE_URL at Postgres for a multi-user/clinical deployment.

Run / register with an MCP host

python -m neuro_mcp     # stdio transport

{
  "mcpServers": {
    "neuro-analysis": {
      "command": "/path/to/envs/eeg-mcp/bin/python",
      "args": ["-m", "neuro_mcp"],
      "env": { "DATABASE_URL": "sqlite:////data/neuro_mcp.db", "BIDS_ROOT": "/data/bids" }
    }
  }
}

neuroii web visualization

Three tools port NEUROII's main views into self-contained interactive HTML files (Plotly, embedded — no server, works offline). Each returns the .html path; interaction runs client-side:

• visualize_timeseries (RawView) — MNE-style stacked channels with page navigation (⏮ ◀ ▶ ⏭), a page-length box, scroll-to-zoom amplitude, and a grid toggle.
• visualize_averaging (EvokedView) — the averaged ERP as stacked channels with a green time cursor + a scalp topomap; a time slider scrubs both, plus a summary sidebar (nave / peak / tmin / tmax).
• visualize_esi (EsiView) — a volumetric source estimate (fsaverage template) rendered to canvas on three orthogonal MRI slices (sagittal/coronal/axial) with a black-blue-white-red activation overlay, crosshair, L/R and MNI-coordinate labels; the cut planes recentre on each frame's peak. Below, the ERP butterfly carries a red current-time cursor and a blue half-peak marker. Controls: time slider, global/frame colormap-scale toggle, and a mask-threshold slider. Faithful port of NEUROII's views; needs epochs (epoch_neuro + set_montage).

visualize_averaging(session_id="s") -> {"out_path": ".../averaging_s.html", ...}

neuroii integration (greenfield)

neuroii integration is not wired yet. The tools define and return the expected REST contract (see neuro_mcp/neuroii/client.py); until NEUROII_API_URL is set they respond {"status": "not_configured", "contract": {…}} so the neuroii app has a fixed target to implement (POST /api/v1/recordings, POST /api/v1/viz-sessions, GET /api/v1/recordings/{id}/annotations).

Testing

python testing/verify.py     # in-memory MCP client, temp SQLite + BIDS, synthetic EEG

Covers rename integrity, the processing core, the full clinician EHR/annotation lifecycle (add → amend → history → void, with audit), and the neuroii stub. For a full-stack run against Postgres, use testing/docker-compose.yml.

Licensing

neuro-mcp is BSD-3-Clause and bundles no third-party source. All required dependencies are permissive (BSD/MIT/Apache-2.0/PSF). Optional extras carry their own terms — psycopg (LGPL-3.0), PySide6 (LGPL-3.0, chosen over GPL PyQt6). Full attribution and compliance notes are in NOTICE.

License

BSD-3-Clause — see LICENSE.