semantic-search-mcp

Semantic code search for AI coding agents. Local embeddings. No API keys. No data leaves your machine.

Your AI agent (opencode, Claude) can grep for exact words - but semantic-search-mcp lets it find code by meaning. Ask "where do we handle authentication?" and it returns auth.controller.ts, login.component.jsx, auth.config.php - even if the word "handle" doesn't appear in any of them.

80MB model. Runs 100% locally. Powered by bge-small-en-v1.5.

Grep vs. Semantic Search

On a 6,900-file codebase:

Quick Start (3 steps)

1. Install

npm install -g semantic-search-mcp

2. Index your project

cd /path/to/your-project
semantic-search-mcp index

The folder you run this from gets indexed. Shows live progress:

████████████████░░░░░░ 70% (5200/7368) - ~120s remaining
██████████████████████ Done! 7368 chunks in 726s.

First run downloads the model (~80MB, one-time) + indexes your code (5-15 min depending on project size). After that, the cache is saved and restarts are instant.

Multiple projects? Run cd /project-a && semantic-search-mcp index, then cd /project-b && semantic-search-mcp index. Each project gets its own cache automatically.

3. Connect your AI agent

Add this to your opencode.json (or opencode.jsonc) in the project root:

{
  "mcp": {
    "semantic-search": {
      "type": "local",
      "command": ["npx", "-y", "semantic-search-mcp"],
      "enabled": true
    }
  }
}

Claude Desktop - add to claude_desktop_config.json:

{
  "mcpServers": {
    "semantic-search": {
      "command": "npx",
      "args": ["-y", "semantic-search-mcp"]
    }
  }
}

Claude Code (CLI) - add .mcp.json to your project root:

{
  "mcpServers": {
    "semantic-search": {
      "command": "npx",
      "args": ["-y", "semantic-search-mcp"]
    }
  }
}

Restart your AI agent. Done. Searches are instant - cache was already built.

Related MCP server: CodeGrok MCP

FAQ

Which folder gets indexed?

The folder you cd into before running semantic-search-mcp index. It's your current working directory. When opencode or Claude starts the MCP server, that same folder gets used automatically.

I have 3 projects. Do I index each one?

Yes. Each project has its own cache:

project-a/.semantic-search/cache/index.json
project-b/.semantic-search/cache/index.json
project-c/.semantic-search/cache/index.json

Where is the cache stored?

{your-project}/.semantic-search/cache/index.json

About 50-100MB per project. Survives PC restarts, Git pulls, everything. It's just files on disk. Only cleared if you run semantic-search-mcp clean.

How do I remove the cache?

semantic-search-mcp clean

Do I need to re-index after code changes?

No. But if you add many new files or want fresh results: semantic-search-mcp clean && semantic-search-mcp index.

What model does it use?

Xenova/bge-small-en-v1.5 by default (80MB, 384-dim, retrieval-optimized). You can switch models via semantic-search-mcp config.

Is my code sent anywhere?

No. Everything runs on your machine - model, embeddings, search. Zero network calls after model download.

CLI Commands

semantic-search-mcp index    # Index current folder (live progress bar)
semantic-search-mcp config   # Interactive TUI to pick extensions, model, thresholds
semantic-search-mcp clean    # Remove index cache
semantic-search-mcp init     # Print opencode/Claude config snippet
semantic-search-mcp           # Start the MCP server (used by AI agents)
semantic-search-mcp --help   # All commands

Configuration

Run semantic-search-mcp config for interactive setup (checkboxes for extensions, searchable model picker, number inputs).

Or create .semantic-search.json in your project root:

{
  "extensions": [".php", ".js", ".jsx", ".ts", ".tsx"],
  "skipDirs": ["node_modules", "vendor", ".git", "dist"],
  "model": "Xenova/bge-small-en-v1.5",
  "chunkThreshold": 300,
  "maxChunksPerFile": 4
}

Or env vars: SEMANTIC_SEARCH_EXTENSIONS=.php,.js, SEMANTIC_SEARCH_MODEL=Xenova/bge-small-en-v1.5

All options

How It Works

1. Scan - walk your project, find code files

2. Extract - split at function/class boundaries (PHP, JS, TS, Python, Go, Rust, Java)

3. Embed - run each chunk through a local ONNX model (384-dim vectors)

4. Cache - save everything to disk

5. Search - embed your query, find closest matches via cosine similarity

License

MIT