Tool Reference

neuro-mcp registers 54 tools on the neuro-analysis FastMCP server, across five areas:

Area	Tools	State
Processing (MNE)	24	In-memory, keyed by session_id
Source Imaging (ESI)	9	Same session, after epoching
Data & EHR	15	Persistent (Postgres/SQLite + BIDS)
neuroii Integration	3	Persistent + external HTTP
Visualization	3	Reads the in-memory session, writes standalone HTML

Recommended pipeline order

This is the order embedded in the server's own instructions string, which any MCP client (including the calling agent) can read directly:

load_neuro -> filter_neuro -> set_montage -> set_reference ->
detect_bad_channels -> run_ica / apply_ica -> find_events -> epoch_neuro ->
compute_psd / compute_erp / time_frequency

# source imaging, after epoching:
fetch_template_head -> compute_forward -> compute_noise_covariance ->
make_inverse_operator -> apply_inverse (or apply_lcmv_beamformer) ->
extract_label_timecourses

Pick a stable session_id (e.g. "default", or one per patient/recording) and reuse it across calls — state (loaded raw data, fitted ICA, epochs, forward/inverse operators, source estimates) threads through that key. Data and EHR tools are independent of this: subjects, recordings, and EHR records persist in the database across sessions and process restarts.

Two state models, two return-key conventions

Processing tools mutate transient in-memory session state and tend to key their status as "status" (e.g. filter_neuro returns {"status": "filtered", ...}). Data/EHR tools mutate the persistent store and key their result as "outcome" (e.g. add_ehr_record returns {"outcome": "created", ...}). There's no shared response envelope across the two families — when calling a tool for the first time, check its actual return shape rather than assuming one convention holds everywhere.

See Examples for complete, verified call sequences.