Hosted
Connect to the Nimbus cloud backend with a token from nimbus-mcp login, or mount the shared HTTP gateway. Your plan's quotas apply.
Open Source · MCP Server
Your AI agent just became a BCI engineer.
nimbus-mcp connects Claude Code, Cursor, and Claude Desktop to Nimbus Studio. Build and validate EEG pipelines, launch experiment campaigns, stream live sessions, and calibrate a personal classifier — all from a terminal conversation.
$ uvx nimbus-mcp login
✓ token saved → ~/.nimbus/credentials.json · plan: pro
$ claude mcp add nimbus -- python -m nimbus_mcp
✓ connected · 58 tools
> benchmark motor imagery on BNCI2014-001,
then sweep 5 configurations
● catalog.leaderboard → experiment.run ×5 → done
mean accuracy 0.71 ± 0.04 · artifacts saved to project
58
MCP tools across the whole workflow
25
runs in a single experiment campaign
3
agent clients supported out of the box
WHAT YOU CAN ASK
58 tools. Infinite ways to ask.
Click any question to see how agents use MCP tools to solve it, or copy the prompt into your agent.
“Which EEG foundation model performs best on my data?”
↳Ranks REVE, LaBraM & EEGNet by accuracy
“Find the fastest pipeline with over 70% accuracy”
↳Sweeps 16 setups to cut compute by 52%
“What is the best published pipeline for motor imagery?”
↳Queries verified MOABB benchmark rankings
“Turn this visual pipeline into standard Python code”
↳Exports clean scikit-learn code with zero lock-in
“Find public datasets with 64 channels and P300”
↳Discovers matching benchmarks with auto-download
“Which EEG foundation model performs best on my data?”
↳Ranks REVE, LaBraM & EEGNet by accuracy
“Find the fastest pipeline with over 70% accuracy”
↳Sweeps 16 setups to cut compute by 52%
“What is the best published pipeline for motor imagery?”
↳Queries verified MOABB benchmark rankings
“Turn this visual pipeline into standard Python code”
↳Exports clean scikit-learn code with zero lock-in
“Find public datasets with 64 channels and P300”
↳Discovers matching benchmarks with auto-download
“Why is my headset signal noisy right now?”
↳Checks impedance & flags poor electrode contact
“Calibrate a personal classifier for my participant”
↳20-cue guided session adapted in <20ms
“Set up a neurofeedback protocol with sham control”
↳Double-blind training with instant pre/post stats
“Stream live EEG directly into my coding agent”
↳Sub-20ms private streaming with auto-watchdog
“Package my recording into standard BIDS format”
↳Generates BIDS sidecars, events & checksums
“Show me which brain states were confused in testing”
↳Generates publication-ready confusion charts
“Why is my headset signal noisy right now?”
↳Checks impedance & flags poor electrode contact
“Calibrate a personal classifier for my participant”
↳20-cue guided session adapted in <20ms
“Set up a neurofeedback protocol with sham control”
↳Double-blind training with instant pre/post stats
“Stream live EEG directly into my coding agent”
↳Sub-20ms private streaming with auto-watchdog
“Package my recording into standard BIDS format”
↳Generates BIDS sidecars, events & checksums
“Show me which brain states were confused in testing”
↳Generates publication-ready confusion charts
One prompt, a full experiment.
you
›Benchmark motor imagery pipelines on BNCI2014-001 and tell me what wins.
● catalog.leaderboard --dataset BNCI2014-001
● catalog.templates --paradigm motor_imagery
● experiment.run --configs 5
✓ aggregated metrics → project
baseline
0.71
csp_ldariemanneegnetxdawnfilter_bank
YOU GET
A ranked leaderboard of 5 pipelines — best holds 0.71 ± 0.04 accuracy, with artifacts and metrics already saved to your project.
How it works
From pip to first pipeline
01 · Install
Or skip installing entirely — uvx runs it straight from PyPI.
02 · Log in
Device-code flow, no secret copy-paste; the desktop app key is picked up automatically.
03 · Connect
Cursor, Claude Desktop, and the hosted gateway — every config is one README away.
Local mode, env keys, and the HTTP gateway are documented in the README.
Runs wherever your backend is
Local
Pair with the Nimbus Studio desktop app backend on 127.0.0.1:8080. The desktop key is detected automatically and file reads stay on your machine.
Setup mode
No credentials yet? The server still boots, and every tool call returns onboarding instructions — your agent can walk itself through setup.
An agent at the keyboard, rules on the wire
Consent-gated hardware
stream.start and calibration.start require an explicit confirmation — your agent has to ask before anything touches a device.
Human in the loop
project.save hands pipelines to Nimbus Studio for review, so nothing ships without you.
Scoped credentials
Device-code login with revocable tokens — 30 days by default, 90 at most. Local-mode keys never leave your machine.
Idle watchdog
Live streams stop themselves after 15 idle minutes, so a forgotten session can't record forever.
Frequently Asked Questions
What is an MCP server?
MCP (Model Context Protocol) is the open standard that lets AI agents call external tools. nimbus-mcp is an MCP server for Nimbus Studio: it exposes 58 tools your agent can call to build, validate, and run BCI pipelines — no UI clicking required.
Which AI clients does it work with?
Claude Code, Cursor, and Claude Desktop connect over stdio; any MCP-compatible client can use the hosted HTTP gateway instead. Ready-made configs for every client are in the README.
Is nimbus-mcp free?
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; calibration workflows and custom data uploads require a Studio Pro plan ($29/month).
Does my EEG data leave my machine?
Not in local mode: the server pairs with the desktop app backend on 127.0.0.1:8080 and file reads stay on your machine. In hosted mode your data goes to the Nimbus cloud backend under your account — encrypted in transit, isolated per account.
Can the agent start my hardware without asking?
No. Live streaming and calibration are confirm-gated: the tool call must carry an explicit confirmation, so your agent has to ask you first. Idle streams stop themselves after 15 minutes.
What can I ask my agent to do?
Benchmark pipelines on 90+ public datasets, sweep up to 25 configurations in one experiment campaign, check live headset quality, run a guided calibration, train your personal classifier, and export everything to BIDS — the examples above show each of these end to end.
Open source, agent-ready.
$pip install nimbus-mcp
Free to use with your Nimbus account: catalog browsing and runs on public datasets work on the free tier. Calibration workflows and custom data uploads require Pro.