ephys-mcp

An MCP server that lets an LLM analyse intracortical (spike-level) brain-computer-interface recordings: signal quality, spike detection, firing rates, and cursor-velocity decoding.

Existing BCI MCP servers target scalp EEG. This one targets the kind of data a high-channel-count implant produces, and defines a read-only adapter contract so a live device backend can be added when a vendor publishes an API.

Research and education software. Not a medical device. Not for clinical use. Not affiliated with or endorsed by Neuralink Corp. or any other implant manufacturer.

Example output

Figures from the built-in synthetic source (32 units, 300 s, seed 2), produced by the server's own plot tools.

plot_psth grouped by reach direction: population rate per direction (mean ± SEM over 166 trials) above a unit-by-time heatmap of change from baseline	plot_decoding after fit_decoder(kind="kalman"): decoded against actual cursor velocity on a held-out 10 s window, R² 0.90 (x) and 0.83 (y)

Status

v0.5. The planned feature set is complete: local NWB files, local broadband WAV recordings, live Lab Streaming Layer streams, streaming from the DANDI Archive, a synthetic motor-cortex source with ground truth, spike detection, quality metrics, ridge and Kalman decoders, trial-aligned PSTHs, spike sorting, cross-session (FALCON-style) evaluation, latent-factor models (GPFA, PCA), probe geometry, and figures. 24 tools. Bug reports and feature requests go to GitHub Issues.

Install and run

Needs uv. No install step: uvx ephys-mcp fetches the package and starts the server on stdio.

Claude Code:

claude mcp add ephys -- uvx ephys-mcp

Claude Desktop (claude_desktop_config.json):

{ "mcpServers": { "ephys": { "command": "uvx", "args": ["ephys-mcp"] } } }

From a checkout, use uv run ephys-mcp instead, or uv --directory /path/to/ephys-mcp run ephys-mcp in the configs above.

HTTP transport

For remote clients or hosted agents, serve streamable HTTP instead of stdio:

EPHYS_MCP_TOKEN='a-long-random-secret' uvx ephys-mcp --http --host 0.0.0.0 --port 8000

Every request must then carry Authorization: Bearer <token>. The server refuses to bind to a non-loopback address without a token, and tokens must be at least 16 characters. Put TLS in front of it (a reverse proxy) before exposing it beyond a private network: the token travels in clear text otherwise. On loopback the token is optional, so ephys-mcp --http alone serves http://127.0.0.1:8000/mcp for local testing.

Then ask, for real data: "Find a small motor cortex dataset on DANDI, open it, and tell me how well hand velocity can be decoded." Or offline: "Open a synthetic session, check signal quality, fit a Kalman decoder and show me a decoded window."

Data sources

Dataset licence and citation come from the archive and are returned by open_session, so the model can attribute the data. Many datasets record only during trials; the server tracks those spans (recorded_fraction) and leaves the gaps out of rates and decoding instead of reading them as silence.

WAV samples carry no physical unit, so amplitudes are reported as ADC counts unless you pass uv_per_count; every amplitude result names its unit. Clips in a folder are separate recordings, so the server says that timing across those channels is not meaningful. Spike times from WAV are threshold crossings, not sorted units.

Reference results, all simple causal linear baselines rather than state of the art:

• MC_Maze_Small (DANDI 000140, 142 units, last 20% held out, 50 ms bins): ridge R² 0.50, Kalman R² 0.34 for hand velocity.
• FALCON H1 (DANDI 000954, human 7-DoF velocity, 176 channels, 20 ms bins, eval_mask): ridge trained on the first held-in day scores R² 0.43 on that day's minival, 0.07 one week later and below zero on the held-out days. That decay is the point of the benchmark; the Kalman filter is unsuitable for this scripted calibration data. FALCON's official test labels are private, so these are not leaderboard scores.

Decoder hyperparameters (ridge strength, the neural lead for Kalman) are chosen by blocked cross-validation inside the training split. Ridge history is 0.5 s of spike counts whatever the bin size.

GPFA is implemented from the paper's equations in numpy and scipy (EM over loadings, offsets, noise and per-factor timescales; no deep-learning dependency), and runs in seconds on a hundred trials. On the simulator, whose true latent is 2-D cursor velocity, it finds two dominant factors that explain velocity with R² 0.95 (PCA: 0.68). On MC_Maze_Small it shows the rotating population trajectory around movement onset that motor cortex is known for. LFADS-class models are out of scope: they need a training run of minutes and a deep-learning stack.

Tools

Resource: ephys://sessions. Prompts: analyze_session, falcon_evaluate.

Tools return summaries, never raw arrays, so results fit in a model's context.

Optional extras

Sorting uses the session's probe geometry, from the file's electrode table or from set_probe_geometry, so contacts under 100 µm apart are sorted jointly. This matters on dense probes: on a simulated 20 µm laminar probe, sorting with the true geometry finds the 12 real units, while treating contacts as independent reports 15, counting the same unit again on neighbouring contacts. Without geometry, channels are placed far apart and treated as isolated electrodes. On the simulator, spykingcircus2 recovers every unit with recall above 0.95.

Brain-area labels come from the NWB electrodes table when present (MC_Maze reports PMd and M1); get_probe lists them per channel so an analysis can be restricted to one area with the units argument. Note that none of the DANDI datasets tried so far stores contact coordinates, so for those set_probe_geometry is the way to supply an array layout.

Plot tools return the PNG inline, so a vision-capable model can read the figure, and also save it under ~/.cache/ephys-mcp/plots (override with EPHYS_MCP_OUTPUT_DIR). Figures use a categorical palette checked for colour-blind separation, with direct labels so identity never rests on colour alone.

Design rules