BCI Features

Previous Next

Device backends

BCI-MCP uses a URI-based device registry. Every backend implements the same Device ABC (bci_mcp.core.device.Device) and registers a URI scheme.

Backend URI prefix Notes
Synthetic synthetic:// No hardware; generates band-mixed EEG
NeuroFocus v4 neurofocus:// Serial (neurofocus://serial/PORT) or BLE (neurofocus://ble/NAME)
BrainFlow brainflow:// OpenBCI Cyton, Ganglion, Muse 2/S, and more
LSL lsl:// Any Lab Streaming Layer-compatible device
Generic serial serial:// One integer sample per line at 115200 baud
Playback playback:// Replay a recorded .npz, .csv, or .edf file

The Device produces Chunk objects: data is (channels, n_samples) float32 in µV, timestamps is (n_samples,) in seconds.

Acquisition architecture

Device.read() → Chunk → Stream → RingBuffer → consumers (Pipeline, Recorder, LSL publisher)

Stream runs in a background thread, calling device.read() at ~4× the chunk rate and feeding a RingBuffer. Consumers call stream.latest(n) to pull the most recent n samples.

Recording and replay

Record to file:

bci-mcp record --device synthetic:// --seconds 60 --out session.npz
bci-mcp record --device synthetic:// --seconds 60 --out session.csv
bci-mcp record --device synthetic:// --seconds 60 --out session.edf   # needs [edf]

Replay:

bci-mcp play session.npz

The PlaybackDevice reads the recording at the original sample rate and loops (optional ?loop=true).

Neurofeedback trainer

bci-mcp neurofeedback --device synthetic:// --metric focus --target 0.7 --seconds 60

NeurofeedbackSession samples the pipeline at 4 Hz, tracks time in-zone (metric ≥ target), streak, and cumulative percentage. The session summary includes time_in_zone_pct, mean_score, and best_streak.

Available metrics: focus, calm, attention, engagement, fatigue, meditation.

Web dashboard

bci-mcp dashboard --device synthetic:// --host 127.0.0.1 --port 8000

FastAPI + uvicorn serve a /state JSON endpoint and a live dashboard at http://127.0.0.1:8000. Requires the [dashboard] install extra.

LSL publisher

The pipeline can publish its output to a Lab Streaming Layer outlet, making it compatible with the broader BCI ecosystem (OpenViBE, BCI2000, BCILAB, etc.). This is wired in via bci_mcp.lsl.publisher.

Adding a device backend

  1. Subclass bci_mcp.core.device.Device.
  2. Implement connect(), start(), read() → Chunk | None, stop(), disconnect().
  3. Register: from bci_mcp.core.registry import register; register("myscheme", factory_fn).

See devices/synthetic.py for the simplest example.

Configuration MCP Integration