Introducing Nimbus MCP: Give Your Agent a Brain
🚀 nimbus-mcp is live. Brain-computer interfaces, for your agent.
Today we're launching Nimbus MCP — an open-source Model Context Protocol server that connects Claude Code, Cursor, and Claude Desktop to Nimbus Studio. Instead of writing BCI plumbing by hand, you describe the experiment in plain words and your coding agent builds it: it picks datasets, assembles the pipeline, trains the classifier, cross-validates, and hands you the results — 63 MCP tools covering the whole workflow.
The 42-second video above shows the whole loop on a real task.
What the agent actually does
In the video, the prompt is:
Benchmark a motor-imagery pipeline on BCI Competition data — show me the confusion matrix.
That single sentence turns into a four-step plan your agent executes step by step:
- load dataset — pull a public BCI Competition recording from the catalog
- spatial filter — configure the CSP-style spatial filtering stage
- train classifier — fit the decoder on the training splits
- cross-validate — score it honestly across folds
The result comes back as real artifacts, not prose: a confusion matrix for BCI Competition IV-2a (subject 1, 67% accuracy in the video run), a power spectrum, and per-fold cross-validation scores. Real datasets, real scores — no screenshot theater.
The same pattern extends across the rest of the toolbox:
- Benchmarking — rank motor-imagery pipelines on 90+ public datasets, with verified MOABB leaderboard lookups
- Optimization — sweep up to 25 configurations in one experiment campaign and surface the cheapest setup above your accuracy floor
- Live hardware — check headset impedance and channel quality mid-session, before you waste a recording
- Calibration — run a guided 20-cue session and train a personal classifier, adapted in under 20 ms
- Open science — package recordings into BIDS format, export pipelines as standalone scikit-learn Python
Three commands to start
Set up is deliberately boring:
pip install nimbus-mcp
nimbus-mcp login
claude mcp add nimbus
No secret copy-paste: nimbus-mcp login uses a device-code flow. Cursor and Claude Desktop configs are one block in .cursor/mcp.json — exactly what the video types out — and the README covers the hosted HTTP gateway for any other MCP-compatible client. If you'd rather not install anything, uvx runs the server straight from PyPI.
Runs wherever your backend is
- Hosted — connect to the Nimbus cloud backend with a login token, or mount the shared HTTP gateway
- Local — pair with the Nimbus Studio desktop backend on 127.0.0.1:8080; the key is detected automatically and file reads never leave your machine
- Setup mode — no credentials yet? The server still boots, and every tool call returns onboarding instructions, so your agent can walk itself through setup
An agent at the keyboard, rules on the wire
Handing an agent tools that touch EEG hardware deserves guardrails, so the dangerous paths are gated by design:
- Consent-gated hardware —
stream.startandcalibration.startrequire explicit confirmation; the agent has to ask before anything touches a device - Human in the loop —
project.savehands pipelines to Nimbus Studio for your review, so nothing ships without you - Scoped credentials — revocable device-code tokens, 30 days by default; local-mode keys never leave your machine
- Idle watchdog — live streams stop themselves after 15 idle minutes
Try it
The server is open source and free to use with a Nimbus account — catalog browsing and runs on public datasets work on the free tier.
- Full walkthrough, examples, and FAQ: nimbusbci.com/mcp
- Source: github.com/nimbusbci/nimbus-mcp
- Install: nimbus-mcp on PyPI
Give your agent a brain. Then ask it for the confusion matrix.
Ready to run real-time Bayesian BCI?
Design and simulate pipelines visually in Nimbus Studio, or install the sklearn-compatible Python SDK for sub-20ms local inference.
Related Articles
Alpha You Can See: Checking Your EEG Signal Before You Trust It
Connect a headset and Studio shows the live signal with a per-channel verdict: 20-bit range with sub-microvolt steps, contact offsets removed before the pipeline, and the eyes-closed alpha rhythm answering on screen at 2.5-3x.
No Install, No Middleware: Connecting EEG Hardware Directly in the Browser
Twelve EEG device types connect straight from a web page: Web Bluetooth for BLE headsets (BrainBit family, Muse, OpenBCI Ganglion), Web Serial for the OpenBCI Cyton dongle. The browser is the driver.