Every harness already has grep and file reads. What is missing is rationale — the decision that produced a function, the invariant it holds, the issue thread that explains the odd branch. rag-rat stores those as typed notes bound to symbols, and every answer carries a confidence label and coverage warnings so a partial index reads as exactly that.
A local Rust index and MCP server over one checkout. It builds a tree-sitter call graph, reads git and GitHub history, and stores typed source-anchored notes, then answers with provenance on every result. Source files stay read-only; it writes only its own SQLite database.
- 47 tools in nine groups, covering search, the call graph, clones, history, index health and repo memories
- semantic_search, symbol_lookup and docs_for_symbol for finding your way in
- find_callers, trace_callees, impact_surface and check_library_usage — what calls this, what it calls, and what breaks if you change it
- find_clones and clones_for_symbol — exact and near-duplicate functions already in the repo
- git_history_for_symbol, git_blame_chunk, commit_search and the papertrail tools — the issues, PRs and reviews behind a line
- memory_create and the memory_* family — typed notes (Invariant, Decision, Risk, BugPattern and more) anchored to a symbol, path or call path, retrievable later with evidence
- dream and dream_review — a maintenance pass over accumulated memories
- Results encoded as TOON rather than JSON, which is denser for the uniform result rows
Install the plugin for Claude Code, Codex or opencode, which registers the server and installs a matching binary; or install the CLI from npm as @rag-rat/bin, or build it with cargo. The registry entry is io.github.cq27-dev/rag-rat over stdio. The server starts dormant until the repo is indexed.
One command — rag-rat mcp