brain0

What it is

brain0 is a Rust-core CLI (npx brain0 up) that runs as a pure, read-only observer over a git repo and its coding-agent transcripts. The summary describes what it produces; the page adds how you run it, what it requires, and the trust limits of its DLP output.

How it works

A Rust core observer reads git history (the facts) and passively ingests Codex / Claude Code transcripts (the declared intents), extracts symbols and AST fingerprints via Tree-sitter, reconciles declared-vs-done to compute drift, and serves a PixiJS graph GUI at http://localhost:8787. Storage is pluggable behind one Storage trait: SQLite + sqlite-vec (local, open core) vs Postgres + pgvector (enterprise). It also exposes an MCP query channel so your agent can pull provenance-aware context by reference.

Preconditions

• Node.js ≥ 20 is the only hard requirement. Repo id is inferred from the git remote; there is a no-git fallback (fs-watcher / checkpoint engine).
• Offline-first with zero egress: works end-to-end with deterministic summaries and local feature-hash embeddings, no models installed.
• Optional local Ollama models improve summaries/embeddings; summaries are content-cached in ~/.cache/brain0/, so rebuilds cost zero model calls.

What to do

npx brain0 up
brain0 today             # morning triage: riskiest first
brain0 report            # accountability report (add --md to share)
brain0 query "why did the parser break"

Optional local models:

ollama pull qwen3:4b
ollama pull qwen3-embedding:0.6b

Low-VRAM (~4 GB) first run:

BRAIN0_SUMMARIZER_MODEL=qwen3:1.7b npx brain0 up

Agent memory via MCP:

claude mcp add brain0 -- brain0 mcp

Model choices are overridable via BRAIN0_SUMMARIZER_{PROVIDER,MODEL,ENDPOINT}, BRAIN0_EMBED_{PROVIDER,MODEL,ENDPOINT,DIM}, BRAIN0_LLM_PROVIDER; a missing model degrades gracefully (embedder falls back to nomic-embed-text, then local).

Caveats

• The DLP read-set is a lower bound, not proof. It derives from a single source — the transcript the agent harness writes (explicit Read calls, Grep results, shell reads like cat .env, subagent transcripts). Treat it as the minimum that may have reached a possibly-remote model, not the maximum.
• Secret detection records paths and secret kinds only, never values.
• MCP-based agent memory is not persistent by nature: "uninstalling brain0 makes it forget again."
• Local models are positioned as strictly optional — they "make it better, never required."