RepoBrain – Repo Intelligence Engine
Give your repo a brain: ChatGPT for your codebase, works across Claude Code, Cursor, Codex, Windsurf, and more.
Evidence shows: README describes permission controls (e.g., RB_ALLOW_MCP requires explicit opt-in), sandbox configuration (SANDBOX_TYPE default local, not isolated), sensitive data handling (retrieval graph redacts common secrets, but full mode may preserve source snippets), dependency security (CI tests exist, but no dependency vulnerability scanning evidence), external effects (MCP client can connect to external servers, requires explicit enablement), rollback (rb init --force can overwrite, but no version rollback mechanism), source attribution (LICENSE and SECURITY.md present, but publisher unverified). Deductions: least privilege not fully demonstrated, user confirmation only partial (e.g., MCP requires explicit opt-in), data flow transparency insufficient, sensitive data handling limited, dependency security evidence weak, external effects control needs more detail, rollback mechanism missing, source attribution based only on file declarations.
Evidence shows: README and code structure are consistent, tests cover doctor command and init force behavior, CI configured for multiple Python versions. Deductions: dependency availability not fully verified (e.g., fallback when requests missing), failure messages partially covered in tests but not comprehensive.
Evidence shows: README identifies target users (developers, AI IDE users) and use cases (codebase Q&A, knowledge base building), capability boundaries described (e.g., rb-refresh requires LLM, rb init does not), trigger precision explained (e.g., rb-ask requires rb-refresh first), environment fit has support matrix (Claude Code, Codex, etc.). Deductions: some scenario descriptions are brief, capability boundaries not fully explicit.
Evidence shows: README structure is clear, installation notes provided (multiple methods), examples (command examples), known limitations (e.g., sandbox not isolated), license (MIT), version info (v0.3.0 upgrade test). Deductions: naming stability has rename history (Antigravity to RepoBrain), version changelog not provided, maintenance responsibility unclear (publisher unverified).
Evidence shows: output usability described (answers with file paths and line numbers), marginal value compared (vs traditional grep), cost-benefit has benchmark data (2.1× speed improvement). Deductions: cost-benefit data not independently verified, and depends on LLM API costs.
Evidence shows: README provides benchmark report link (artifacts/benchmark-2026-05-09/REPORT.md), CI tests exist, but no independent verification. Deductions: claim traceability limited, cross-source corroboration insufficient, fact-inference separation not explicit.
- Publisher identity unverified; treat source with caution.
- Sandbox is not isolated by default; not suitable for untrusted code.
- MCP external server connections require explicit opt-in, but permissions need attention.
- Benchmark data not independently verified; performance claims may be overstated.
- No dependency vulnerability scanning evidence provided.
What does this agent do, and when should you use it?
RepoBrain (formerly Antigravity Workspace Template) is a cross-IDE repository knowledge engine that uses a multi-agent cluster to autonomously read code, generate knowledge artifacts, and answer questions grounded in actual source with file paths and line numbers. It ships as two Python packages (cli and engine) plus native plugins for Claude Code and Codex CLI. Core components include Router, ModuleAgent, GitAgent, and a multi-agent refresh pipeline. The engine creates shared knowledge files in .repobrain/ that any IDE can read, and it offers an optional MCP server. Installation via plugin marketplace or pip; requires Python 3.10+ and an OpenAI-compatible API key. Main commands: rb-setup, rb-refresh, rb-ask, rb-init.
RepoBrain provides four primary slash commands: rb-setup interactively configures LLM provider and writes .env; rb-refresh deploys a multi-agent cluster that scans the codebase, generates conventions.md, structure.md, knowledge_graph.json, agents/*.md docs, and map.md routing index; rb-ask routes questions to the appropriate ModuleAgent and returns grounded answers with file paths and line numbers; rb-init scaffolds a new multi-agent repository. The engine includes a GitAgent for analyzing git history, and an optional MCP server (rb-mcp) exposing ask_project and refresh_project tools. Installation via Claude Code plugin marketplace or pipx for Codex.
- A developer in Claude Code runs /repobrain:rb-ask "How does auth work?" to get a grounded answer with file references instead of manually searching.
- A team onboarding a new codebase runs rb-refresh to auto-generate module knowledge docs, accelerating ramping up.
- A Codex CLI user asks /rb-ask "Where is X defined?" to quickly locate symbols and call chains.
- A project maintainer runs /repobrain:rb-refresh quick after code changes to incrementally update the knowledge base.
- A security auditor uses rb-ask to check for vulnerabilities such as missing authorization, based on code evidence.
What are this agent's strengths and limitations?
- Cross-IDE compatibility: the same .repobrain/ knowledge files are used by Claude Code, Cursor, Codex, Windsurf, etc., avoiding per-IDE configuration.
- Automated knowledge base generation: rb-refresh uses multi-agent cluster to generate module docs, reducing manual documentation effort.
- High factual accuracy: benchmark shows 99% accuracy on factual lookups and is 2.1× faster than Codex CLI on factual questions.
- Language-agnostic: module detection based on directory structure, works with any programming language without AST parsing.
- Requires LLM API key: rb-refresh and rb-ask require an OpenAI-compatible API key; local mode is for personal use only.
- Initial refresh is time-consuming and consumes API calls; large codebases may take minutes and incur cost.
- Quality depends on chosen LLM provider, introducing potential vendor lock-in.
- Only Claude Code and Codex have native plugin support; other IDEs rely on file integration or MCP, which may lack full functionality.
How do you install or deploy this agent?
Claude Code plugin: in Claude Code, run /plugin marketplace add study8677/repobrain then /plugin install repobrain@repobrain. Codex CLI manual install: pipx install "git+https://github.com/study8677/repobrain.git#subdirectory=engine" and pipx inject --force --include-apps repobrain-engine "git+https://github.com/study8677/repobrain.git#subdirectory=cli", then add plugin marketplace. Requires Python 3.10+ and an OpenAI-compatible API key.
How do you use this agent?
After installation, in your project run /repobrain:rb-setup to choose LLM provider and save .env, then /repobrain:rb-refresh to build the knowledge base, and finally /repobrain:rb-ask "How does auth work?". For Codex, use /rb-setup, /rb-refresh, /rb-ask. For manual install, use rb-refresh --workspace . and rb-ask "question" --workspace ..
FAQ
How is RepoBrain different from Codex CLI's built-in code search?
Do I need to run rb-refresh for every project?
/repobrain:rb-refresh quick for incremental updates.