A research lab's worth of agents, one PI in charge

What it does

BrainPilot is an open-source agentic research workspace for neuroscience, built in TypeScript. A Principal Investigator agent talks to you, plans the work, and delegates to specialist agents covering literature review, experiment design, analysis, writing, and auditing. Every session is recorded as a Graph of Trace, so intermediate actions, evidence, and claims stay inspectable rather than vanishing into a prompt.

The interesting bit

The Auditor agent is the standout: it reviews claims, evidence chains, citations, and hallucination risks — an AI checking the other AIs’ homework. The case studies are refreshingly honest about statistical limits, reporting non-significant p-values (0.058–0.243) alongside the wins.

Key highlights

• PI agent coordinates librarian, experimentalist, engineer, writer, and auditor
• Graph of Trace makes the full research process inspectable, with workspace recovery
• Real case studies on Allen Neuropixels, BCI Competition, and fMRI pain data
• Pluggable providers (Anthropic, OpenAI-compatible, Azure) plus MCP tool support
• Runs locally as a single-user process; mock mode works without an API key

Caveats

• Local (non-Docker) mode has no container isolation — agents read and write directly on your machine
• Single-user data root; multi-user hosted deployments need separate data directories per user
• AGPL v3 license, which may matter for some institutional uses

Verdict

Worth a look if you do neuroscience research and want an agent system that treats auditability as a feature, not an afterthought. If you’re outside brain science or allergic to agent orchestration, the domain packaging won’t convert you.