Semantic Knowledge Retrieval with Claude Code Search MCP

Install and run ChromaDB locally using Docker. This will serve as the storage for your code embeddings, allowing for fast, semantic-based retrieval without sending code to a third-party vector provider.

docker run -d -p 8000:8000 chromadb/chroma

Watch out: Ensure the Docker volume is persisted so your index isn't lost when the container restarts.

Run the MCP indexer script over your source code. The script chunks your files, generates vectors using the OpenAI or Anthropic embedding models, and stores them in ChromaDB.

npx mcp-code-indexer index --path=./src --db=http://localhost:8000

Watch out: Add your build and dependency folders (like node_modules) to the ignore list to avoid polluting the search results.

Define the MCP server in your Claude Desktop configuration. This bridge allows the Claude app to communicate with your local ChromaDB instance and use it as a tool.

{
  "mcpServers": {
    "code-search": {
      "command": "node",
      "args": ["server.js", "--db", "http://localhost:8000"]
    }
  }
}

Watch out: Ensure the path to the server script is absolute to avoid configuration errors during Claude's startup.

Restart Claude Desktop and verify that the 'Code Search' tool appears in the active plugins list. This confirms that the MCP handshake was successful and Claude can now access your index.

Watch out: If the tool doesn't appear, check the Claude logs for JSON parsing errors in your config file.

Start asking Claude questions about your codebase in natural language. Instead of keywords, Claude uses the 'code-search' tool to find logic related to your query's meaning.

Watch out: If results are imprecise, consider re-indexing with smaller chunk sizes (e.g., 500 characters) for better granularity.