Repository navigation
braintrust builds AI personas from what people are actually publishing, and rebuilds them as they publish more. Static personas go stale; a prompt that says respond like X reflects X as they were when you wrote it. braintrust keeps each one current, so when you ask your council a question you get what they think now — dated, and cited back to what they actually published.
Nothing here ships on npm. braintrust is a repo you clone, configure and deploy: a web service and a daily job from one codebase, against your own Postgres and your own two model endpoints. This release is the point the design stopped moving.
What runs end to end
- Ingestion — Substack, YouTube, any blog that declares a feed or a sitemap, and Bluesky a day at a time. Registration prices the work and shows you the plan before it fetches anything.
- The compiler — each item read exactly once, notes kept, four core layers plus positions, revision detection, and a publish gate that keeps yesterday's persona answering rather than shipping a bad rebuild.
- Daily refresh — one cycle, three triggers: the clock, an AI-callable refresh, and following someone.
- The MCP server — all six tools live, over HTTPS with OB1's
?key=convention. - An eval harness —
npm run evalscores the model that reads your corpus, on your corpus. Nothing judges with a model; every measure is a count. - A Hermes agent per Person — one profile each, with a deliberately thin
SOUL.mdso the persona arrives live rather than frozen into a system prompt. Seehermes/.
What it refuses to do
Guess, evade, publish something it cannot defend, or claim a complete corpus it does not have. Paywalled content is never ingested; quotes it cannot find in the source are dropped rather than stored; every persona is named braintrust model of X, never the bare name.
Not in 1.0
- Council mode — one question, every persona answers.
- Revision judgement at scale. The mechanism ships and is tested; the tuning against a real model across a large corpus has not been done.
Getting started
Clone, cp .env.example .env, paste schema.sql into your Supabase SQL editor, deploy twice — npm start and a daily npm run job — then add your first council member through your AI client. Only a human can follow someone, so an AI can refresh a persona but never introduce one. Full instructions in the README, and the reasoning behind every choice in docs/design/.
Built on Open Brain (OB1) by Nate B. Jones. An independent extension, not affiliated with or endorsed by him.