Summary

It tells you when not to trust it, which matters more than the speed numbers.

A query against a repository that is not ready comes back as `index_unavailable` with a safe fallback of native search and a remediation to follow — so a stale index degrades into ordinary grep rather than into confident wrong answers. Check readiness with `get_status` before you rely on results, and note that only the tracked default branch is indexed automatically.

What it is

A local-first search index over your repositories, exposed to coding assistants over MCP. Indexing happens on your machine and the code never leaves it; the assistant searches instead of reading files, which is both faster and cheaper in tokens.

What you get

`symbol_lookup` — exact class or function lookup, benchmarked under 100ms on indexed repos
`search_code` — pattern, keyword or semantic search, benchmarked under 500ms
`get_status` — repository readiness, so you know whether to trust an indexed answer
`reindex` — rebuild the index on demand after a large change
One server across many unrelated repositories, with one registered worktree per repository
Optional semantic search over embeddings, and optional index sharing through GitHub Artifacts with sensitive files filtered on export

Requirements

Run the published container image, or native Python with `uv sync --locked`. Basic search needs no API key; semantic search needs `SEMANTIC_SEARCH_ENABLED` plus a `VOYAGE_API_KEY` or a local endpoint. `MCP_ALLOWED_ROOTS` bounds which paths the tools may touch, `MCP_CLIENT_SECRET` guards the local stdio handshake, and `MCP_AUTO_INDEX` can be set to false on very large repositories so you index on demand.

Setup effort

One command plus a key — docker run -it -v $(pwd):/workspace ghcr.io/consiliency/code-index-mcp:v1.4.0, then supply credentials

Code-Index-MCP