Brain0-ai/brain0

The black box for AI-written code. Passive decision graph linking every commit to the agent prompts behind it: drift detection, DLP audit of what agents read, evidence-driven risk, MCP memory for coding agents, signed provenance attestations. One command, offline by default.

What it solves

Coding agents now write most of the diff, but git only shows what changed, not why. brain0 answers the questions your repo can't: which prompt wrote this bug, what the agent actually read, and how trustworthy that change is. It also reconciles what agents claim to have changed with what git records (drift), detects sensitive reads like .env or private keys, and assigns an evidence-driven risk score that learns from past incidents.

How it works

brain0 is a pure observer. It reads git history (the facts) and the transcripts coding agents (Codex, Claude Code) already write to disk (the declared intents). It builds a decision graph from repo down to module, file, and symbol — each node carrying dated history, drift, and risk. It never modifies your code or git history. Locally, it uses a SQLite index with embeddings; optionally you can plug in Postgres or your own LLM for richer summaries and chat. A GUI lets you explore the graph, and an MCP interface lets coding agents query context, blame, debug, and audit before touching code.

Who it’s for

Developers and teams who use AI coding agents and want accountability: understanding which prompt caused a change, ensuring agents didn't leak secrets, and spotting high-risk patterns. It's also valuable for security-conscious teams and anyone running local-first, offline tooling.

Highlights

  • Passive observation: no hooks, no agent cooperation, no code changes — just reads git and transcripts.
  • Drift detection: reconciles declared vs. done to flag unstaged or undeclared edits.
  • DLP for agent context: records which files sessions read (paths and secret kinds only) and flags sensitive reads that reached a remote model.
  • Risk score that learns: a-priori (blast radius, churn) fused with a-posteriori (reverts, fixes) evidence, surfacing 'gold signals' for dangerous patterns.
  • Local-first and offline by default: no keys, local embeddings, zero egress; optional remote models are opt-in.
  • MCP integration: gives coding agents a query channel for context, blame, debug, and audit before they make changes.
  • Tamper-evident: append-only graph, pinned transcripts, crypto-shred purge, signed attestations.
  • Open core: Apache-2.0, fully functional, with an enterprise option for Postgres and multi-tenancy.

Related

  • Project
    Ar9av/obsidian-wiki

    A system that turns an Obsidian vault into a persistent, agent-readable digital brain, allowing AI agents to store, compile, and retrieve knowledge across sessions without relying on external databases.

  • Project
    vercel-labs/lat.md

    A knowledge graph for codebases written in markdown that links documentation to source code, helping AI agents and humans maintain and navigate domain context.

  • Dispatch
    Version Control for AI Agents: Beyond the Git Commit

    An exploration of Regent VCS, a tool designed to track AI agent intent and session history, and the debate over whether agents need their own version control or can simply use Git.

  • Project
    alibaba/open-code-review

    An AI-powered code review CLI tool that combines deterministic engineering with an AI agent to provide high-precision, token-efficient reviews of Git diffs and full files.

  • Project
    tirth8205/code-review-graph

    code‑review‑graph is a Python package that parses a repository with Tree‑sitter, stores the AST‑derived nodes/edges in a SQLite graph, and serves a minimal, blast‑radius slice of that graph to any MCP‑compatible AI coding assistant. It reduces per‑question token usage by ~63×, supports 30+ languages (including notebooks), offers incremental updates, framework‑aware edges for PHP/Laravel and Java/Spring, a GitHub Action for CI risk‑scored reviews, and optional semantic embeddings. Benchmarks show large token savings and an average F1 of 0.69 for impact analysis. The tool is aimed at developers and teams using LLM‑based code reviewers who need precise, low‑overhead context.