It gives an AI agent one interface over the whole clinical/research EEG workflow: signal processing and source imaging (via MNE-Python), a persistent dataset + EHR store (Postgres + BIDS), and NeuroII web visualization.

Concept

flowchart LR
    Clinician(["🩺 Clinician"])
    Researcher(["🔬 Researcher"])
    Agent[["🤖 AI Agent"]]
    Server(("neuro-mcp<br/>FastMCP · 54 tools"))

    Clinician -- talks to --> Agent
    Researcher -- talks to --> Agent
    Agent -- MCP --> Server

    Server --> Processing["Processing &amp; Source Imaging<br/>MNE-Python + ESI"]
    Server --> Data["Data &amp; EHR Store<br/>Postgres + BIDS<br/>versioned &amp; audited"]
    Server --> NeuroII["NeuroII<br/>Web Visualization"]

    classDef proc fill:#4f8cff,stroke:#2f5fbf,color:#fff
    classDef data fill:#2fb380,stroke:#1c7a55,color:#fff
    classDef viz fill:#b06fe0,stroke:#7c3fae,color:#fff
    class Processing proc
    class Data data
    class NeuroII viz

A clinician or researcher never calls a tool directly — they talk to an agent in plain English, and the agent drives neuro-mcp's 54 tools underneath. See the Tutorial for what that actually looks like end to end.

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_

... View full README on GitHub