repodoc

Auto-documenting repo tool with local RAG retrieval.

• Generates "what-exists" reference docs for each source file (LLM, on push only)
• Maintains them incrementally — a SHA-256 manifest means only changed files are re-documented/re-embedded
• Serves everything through a local hybrid (vector + BM25) index:
  • an MCP server so Claude Code retrieves only the relevant slices (token savings)
  • a MkDocs Material site for humans

All retrieval is local and free (sentence-transformers + LanceDB). The only API cost is generation, at push time — and even that can be moved to a local model via Ollama.

Install

cd c:\apps\repodoc
uv venv --python 3.12
uv pip install -e ".[index,mcp,site]"

The generator needs ANTHROPIC_API_KEY (or [generator] provider = "ollama" in config).

Usage

repodoc init <repo>              # one-time: writes <repo>/.repodoc/config.toml + manifest
repodoc status  --repo <repo>    # what changed since last run
repodoc sync    --repo <repo>    # generate + index changed files (what CI calls)
repodoc query   --repo <repo> "how does the needs system work?"
repodoc serve-mcp --repo <repo>  # stdio MCP server for Claude Code
repodoc build-site --repo <repo> # MkDocs site -> <repo>/.repodoc/site
repodoc reindex --full --repo <repo>  # after changing the embedding model

Useful flags: repodoc generate --dry-run (no LLM calls), --limit N (cap spend per run).

Claude Code integration

.mcp.json in your workspace:

{
  "mcpServers": {
    "gps3-docs": {
      "command": "c:/apps/repodoc/.venv/Scripts/repodoc.exe",
      "args": ["serve-mcp", "--repo", "c:/apps/gps3"]
    }
  }
}

Tools exposed: search_docs(query, k, doc_class), get_doc(path), list_docs().

Keeping docs current on push

• CI (preferred): see gps3/.github/workflows/repodoc.yml — runs repodoc sync on push to the default branch and commits regenerated docs. Needs the repodoc repo on GitHub + ANTHROPIC_API_KEY secret.
• Local hook (opt-in): .git/hooks/pre-push running repodoc sync --repo .. Note this calls the LLM (costs money) on each push — CI with caching is usually better.

Model upgrades

Both model slots live in <repo>/.repodoc/config.toml:

Slot	Change	Consequence
[generator]	e.g. anthropic → ollama + big local model	none — next generate uses it
[embeddings]	e.g. nomic-embed → Qwen3-Embedding 8B	auto-detected → one-time full re-index

Provenance: every indexed chunk is tagged source, generated-reference, or planning, and retrieval surfaces the tag so agents treat planning docs as intent, not ground truth.