turbovec-mcp

Local semantic code search over MCP, backed by turbovec (Google's TurboQuant: ~16x vector compression, fast on Apple Silicon). Bring your own OpenAI-compatible embeddings endpoint - nothing leaves your machine.

code -> chunker -> [your embeddings endpoint] -> turbovec index -> search

• embedder: any OpenAI-compatible /v1/embeddings (a llama.cpp server, etc). Dimension is auto-detected.

• store: turbovec IdMapIndex, persisted per-repo under .turbovec/.

• orchestrator: this MCP server (chunk, embed, index, search).

Install / run

uvx turbovec-mcp        # run directly, no install (recommended)
# or
pipx install turbovec-mcp

Requires a running embeddings endpoint. Example with llama.cpp:

llama-server -m nomic-embed-text-v1.5.Q8_0.gguf --embedding --pooling mean --port 8081

Related MCP server: Claude Context Local

Configure (environment)

Register with an MCP client

opencode (opencode.jsonc):

"mcp": {
  "turbovec": {
    "type": "local",
    "command": ["uvx", "turbovec-mcp"],
    "environment": {
      "TURBOVEC_EMBED_ENDPOINT": "http://127.0.0.1:8081/v1",
      "TURBOVEC_EMBED_MODEL": "nomic-embed-text-v1.5.Q8_0.gguf",
      "TURBOVEC_QUERY_PREFIX": "search_query: ",
      "TURBOVEC_DOC_PREFIX": "search_document: "
    },
    "enabled": true
  }
}

Claude Code:

claude mcp add turbovec -- uvx turbovec-mcp

Tools

• tv_index(path=".") - (re)build the index for a repo. Run once before searching, and after large changes.

• tv_search(query, k=10, path=".", compact=False) - semantic search; returns path, line range, score, and (unless compact) the chunk text.

• tv_status(path=".") - whether a repo is indexed + basic stats.

The index lives in <repo>/.turbovec/ - add it to .gitignore.

License

BSD-3-Clause