grepsense

Two-modal code search for AI coding agents. grepsense gives your agent two tools over MCP:

• code_search — fast lexical search (regex/literal, repo:/file:/lang: filters) via Zoekt, the trigram engine behind Sourcegraph. Sub-10ms.

• semantic_code_search — meaning-based search via local embeddings (ChromaDB + all-MiniLM-L6-v2). Find code by what it does when you don't know the exact name.

It's a single Python MCP server (the embedding model loads once and stays warm) in front of a self-hosted engine. Drop it next to any repo, docker compose up, and point your agent at the URL.

Status: early. Works end-to-end; APIs and layout may still shift.

Quick start (Docker Compose — recommended)

git clone https://github.com/alaiba/grepsense && cd grepsense
cp .env.example .env
#   edit .env → set GREPSENSE_TARGET=/absolute/path/to/your/project
docker compose up -d        # starts Zoekt + ChromaDB + embedder, builds the
                            # index, and serves the MCP server on :8765

Then register the server with your agent (one time):

# Claude Code
claude mcp add --transport http grepsense http://localhost:8765/mcp

The same running stack can serve many clients — run it on a shared host and point everyone's agent at it. The MCP server only needs to reach Zoekt + ChromaDB; only the indexer/embedder mount your source. For Kubernetes/Helm deployments, the chart under charts/grepsense enables health probes by default via probes.enabled, and exposes component-level overrides for mcp.probes, chromadb.probes, zoektWeb.probes, zoektIndexer.probes, and embedder.probes. MCP exposes lightweight HTTP health probes on the same port:

• GET /healthz returns 200 when the MCP process is alive.

• GET /readyz returns 200 only when MCP can reach Zoekt and ChromaDB.

Related MCP server: Code Search MCP

Alternative: native install (no Docker)

pipx install grepsense
cd /path/to/your/project
grepsense embed                      # incremental embed (needs a running ChromaDB)
grepsense embed --full               # force full re-embed of all repos
grepsense embed --reset              # wipe collections and re-baseline
grepsense status                     # per-repo embed state and collection size
grepsense serve --transport http     # or stdio
claude mcp add grepsense -- grepsense serve   # stdio registration

(Native mode still needs a Zoekt webserver and a ChromaDB for the two backends; Compose is the batteries-included path.)

How it works

your code ──► zoekt-indexer ──► zoekt-web (:6070) ─┐
          └─► embedder ──────► chromadb (:8000) ───┤
                                                   ▼
                                   grepsense MCP server (:8765)
                                   code_search · semantic_code_search
                                   stdio · streamable-HTTP

See docs/architecture.md for the design and docs/install.md for the full install/operations guide.

Configuration

Everything is env-driven (see .env.example); a grepsense.yaml can override defaults (include/exclude globs, chunking, embedding model). Key vars:

License

MIT © Vasile Alaiba