Code Search MCP Server

An MCP (Model Context Protocol) server for semantic code search using vector embeddings. Index your code folders and perform semantic searches to find code by meaning, not just keywords.

Features

• Multi-folder support: Index multiple code folders simultaneously
• Semantic search: Find code by meaning, not just keywords
• Continuous file watching: Automatically re-indexes files when they change
• Multiple embedding providers: OpenAI, Ollama, Gemini, Mistral, Bedrock, OpenRouter
• Admin UI: Web dashboard to monitor queries and indexing progress
• 35+ language support: JavaScript, TypeScript, Python, Rust, Go, and more
• Embedding cache: SQLite-based caching to avoid redundant API calls
• LSP enrichment: Type signatures and documentation via Language Server Protocol
• Hierarchical context: Parent class/module context included in embeddings

Prerequisites

Required:

• Node.js 18+
• Qdrant vector database (local or cloud)
• An embedding provider (Ollama for local, or API keys for cloud providers)

Optional (for LSP enrichment):

# TypeScript/JavaScript LSP (improves search quality for TS/JS/TSX/JSX)
npm install -g typescript-language-server typescript

# C# LSP (improves search quality for .cs files)
dotnet tool install --global csharp-ls

Installation

# Clone the repository
git clone https://github.com/amolchanov/mcp-code-search.git
cd mcp-code-search

# Install dependencies
npm install

# Download tree-sitter WASM files
npm run download-wasm

# Build the project
npm run build

Running the Server

The server can run in two modes:

Mode 1: Stdio (for CLI integration)

Used when integrating with Claude CLI, Copilot, or other MCP clients:

# Run directly (waits for JSON-RPC input on stdin)
node dist/index.js

# Or with auto-indexing of specific folders
node dist/index.js --index /path/to/project1 --index /path/to/project2

In stdio mode, the server communicates via stdin/stdout using the MCP protocol. This is what Claude CLI uses when you add it as an MCP server.

Mode 2: SSE with Admin UI (for web dashboard)

Used for the web-based admin interface:

# Start with default port 3100
node dist/index.js --sse

# Or specify a custom port
node dist/index.js --sse --port 8080

# With system tray icon (Windows/macOS)
node dist/index.js --sse --tray

# With auto-indexing
node dist/index.js --sse --index /path/to/project

Then open: http://localhost:3100/admin

Mode 3: Running with PM2 (Auto-Restart on Crash)

For production use, run the server with PM2 for automatic restart if it crashes:

# Install PM2 globally (one-time)
npm install -g pm2

# Start the server with PM2
npm run pm2:start

# View logs
npm run pm2:logs

# Check status
npm run pm2:status

# Stop the server
npm run pm2:stop

# Restart the server
npm run pm2:restart

PM2 will automatically restart the server if it crashes (up to 10 times). Logs are saved to logs/output.log and logs/error.log.

Command Line Options

Option	Description
--sse	Run in SSE mode with HTTP server and Admin UI
--port <number>	HTTP port for SSE mode (default: 3100)
--tray	Show system tray icon (SSE mode only)
--index <path>	Auto-index a folder on startup (can be repeated)

Architecture: Single Server Model

When multiple MCP clients (e.g., multiple Copilot CLI windows) connect simultaneously, the server uses a client-server architecture to prevent database corruption:

┌─────────────────────────────────────────────────┐
│          SSE Server (Daemon)                    │
│  • Handles ALL indexing and database writes     │
│  • HTTP API at localhost:3100                   │
│  • Admin UI for monitoring                      │
└─────────────────────────────────────────────────┘
                     ↑ HTTP
         ┌───────────┼───────────┐
    ┌────┴────┐ ┌────┴────┐ ┌───┴─────┐
    │  stdio  │ │  stdio  │ │  stdio  │
    │ client  │ │ client  │ │ client  │
    └─────────┘ └─────────┘ └─────────┘
         ↑           ↑           ↑
    Copilot 1   Copilot 2   Copilot 3

How it works:

• Each stdio instance (spawned by Copilot/Claude CLI) is a lightweight client
• Clients proxy all tool calls to the central SSE server via HTTP
• The SSE server auto-starts if not running when a client connects
• All database writes (SQLite caches, Qdrant) go through the single server
• This prevents corruption from concurrent writes

Key benefits:

• No database locking issues with multiple CLI windows
• Single indexing process per repository
• Shared embedding cache across all clients
• Consistent state visible in Admin UI

---

Setting Up Prerequisites

Installing Docker (Required for Qdrant)

If you don't have Docker installed:

Windows/macOS:

1. Download Docker Desktop from docker.com
2. Install and start Docker Desktop
3. Verify installation: docker --version

Linux:

# Ubuntu/Debian
curl -fsSL https://get.docker.com -o get-docker.sh
sudo sh get-docker.sh
sudo usermod -aG docker $USER  # Add your user to docker group
# Log out and back in for group changes to take effect

Running Qdrant (Required)

Qdrant is the vector database that stores code embeddings. Choose one option:

Option A: Docker (Recommended)

Start Qdrant container:

# Run in foreground (stops when terminal closes)
docker run -p 6333:6333 qdrant/qdrant

# OR run in background (keeps running)
docker run -d -p 6333:6333 -v qdrant_storage:/qdrant/storage --name qdrant qdrant/qdrant

# Check if running
docker ps | grep qdrant

# Stop the container
docker stop qdrant

# Restart the container
docker start qdrant

# View logs
docker logs qdrant

Verify Qdrant is running:

curl http://localhost:6333/collections
# Should return: {"result":{"collections":[]}, ...}

Option B: Local Installation

Follow the Qdrant installation guide for native installation.

Running Ollama (Optional - for Local Embeddings)

Ollama provides free local embeddings without API costs. Alternatively, you can use OpenAI, Gemini, Mistral, or Bedrock.

Option A: Native Installation (Recommended)

Download and Install:

• Windows/macOS/Linux: Download from ollama.ai
• Follow the installer instructions

Pull the embedding model:

# Recommended model for code (8192 token context)
ollama pull nomic-embed-text

# Verify model is downloaded
ollama list

# Test the model
ollama run nomic-embed-text

Option B: Docker

# Run Ollama in Docker
docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama

# Pull the embedding model
docker exec ollama ollama pull nomic-embed-text

# Verify
docker exec ollama ollama list

Start/Stop Ollama:

# Native installation - Ollama runs as a service automatically
# Check status
curl http://localhost:11434/api/tags

# Docker
docker start ollama
docker stop ollama

Verify Everything is Running

Before starting the MCP server, check:

# Check Qdrant
curl http://localhost:6333/collections

# Check Ollama (if using)
curl http://localhost:11434/api/tags

---

Integration with Claude Code CLI

Step 1: Add MCP Server Configuration

Recommended: Use the Claude CLI command:

# Add to current project (creates .mcp.json in project root)
claude mcp add code-search -s project -- node /path/to/code-search-mcp/dist/index.js

# Or add globally for all projects (user-level config)
claude mcp add code-search -s user -- node /path/to/code-search-mcp/dist/index.js

Windows example:

claude mcp add code-search -s project -- node C:/repos/code-search-mcp/dist/index.js

Verify the server is connected:

claude mcp list
# Should show: code-search: node ... - ✓ Connected

Alternative: Manual JSON configuration

Add to .mcp.json in your project root:

{
  "mcpServers": {
    "code-search": {
      "type": "stdio",
      "command": "node",
      "args": ["/path/to/code-search-mcp/dist/index.js"],
      "env": {}
    }
  }
}

Step 2: Add Instructions for Claude Code

Option A: Use the MCP tool (recommended)

Once connected, ask Claude to run:

Use the get_instructions tool to get the code-search instructions, then add them to my CLAUDE.md file.

Or via CLI:

# The get_instructions tool returns markdown content for CLAUDE.md
# Format options: "claude" (default), "copilot", "generic"

Option B: Copy manually

# Copy to global instructions
cp examples/CLAUDE.md ~/.claude/CLAUDE.md

# Or append to existing instructions
cat examples/CLAUDE.md >> ~/.claude/CLAUDE.md

See examples/CLAUDE.md for the full instructions template.

Key points:

• Prefer mcp__code-search__search over built-in Grep and Glob for finding code
• Use natural language queries: "authentication middleware that validates JWT tokens"
• Fall back to Grep only for exact string/regex matches

---

Integration with GitHub Copilot CLI

Step 1: Add MCP Server Configuration

You can configure the server globally (for all repos) or per-repository:

Option A: Global Configuration (All Repositories)

Add this to your global Copilot CLI MCP settings file:

Location: ~/.copilot/mcp-config.json (create if it doesn't exist)

Windows: C:\Users\<username>\.copilot\mcp-config.json

{
  "mcpServers": {
    "code-search": {
      "type": "local",
      "command": "node",
      "args": ["/path/to/code-search-mcp/dist/index.js"],
      "tools": ["*"]
    }
  }
}

Windows example:

{
  "mcpServers": {
    "code-search": {
      "type": "local",
      "command": "node",
      "args": ["c:/repos/mcp-code-search/dist/index.js"],
      "tools": ["*"]
    }
  }
}

Option B: Repository-Specific Configuration

Add .copilot/mcp-config.json to your repository root:

# In your project repository
mkdir -p .copilot
touch .copilot/mcp-config.json

Add the same configuration:

{
  "mcpServers": {
    "code-search": {
      "type": "local",
      "command": "node",
      "args": ["/path/to/code-search-mcp/dist/index.js"],
      "tools": ["*"]
    }
  }
}

Note:

• Repository-specific config takes precedence over global config
• Don't commit .copilot/mcp-config.json to version control (add to .gitignore)
• If you already have MCP servers configured, add the code-search entry to the existing mcpServers object

Step 2: Add Instructions for Copilot

Copy the example instructions to your Copilot instructions file:

See examples/COPILOT.md for the full instructions template.

Key points:

• Prefer semantic search over grep for finding code
• Use descriptive natural language queries
• Search before implementing to find existing patterns

---

Admin UI Features

The Admin UI is available when running in SSE mode (see Running the Server):

node dist/index.js --sse
# Then open: http://localhost:3100/admin

Folders Tab:

• View all indexed folders with status
• Add new folders via path input
• Remove folders (deletes indexed data)
• Reindex folders (useful after upgrading to get new features)

Queries Tab:

• Query history with expandable results
• Timeline or per-folder view
• Clear query logs

Ingestion Tab:

• Real-time indexing progress
• File counts and error tracking

Services Tab:

• Ollama status and controls (start/stop)
• LSP enrichment status

Settings Tab:

• Embedding cache statistics
• Cache warming controls
• File watcher configuration

---

Configuration

Use the configure tool to set up the server after connecting.

Ollama (Local, Free)

{
  "qdrantUrl": "http://localhost:6333",
  "embedderProvider": "ollama",
  "ollamaBaseUrl": "http://localhost:11434",
  "modelId": "nomic-embed-text:latest"
}

OpenAI

{
  "embedderProvider": "openai",
  "openAiApiKey": "sk-...",
  "modelId": "text-embedding-3-small"
}

Other Providers

Supported: gemini, mistral, bedrock, openrouter, openai-compatible

Per-Folder Embedding Models

You can configure different embedding models for different folders via the Admin UI. This is useful when:

• Some projects need higher quality embeddings (use larger models)
• Some projects need faster indexing (use smaller models)
• Testing different models for comparison

Supported models and their context sizes:

Model	Provider	Context (tokens)
nomic-embed-text	Ollama	8192
mxbai-embed-large	Ollama	512
text-embedding-3-small	OpenAI	8191
text-embedding-3-large	OpenAI	8191
text-embedding-ada-002	OpenAI	8191
mistral-embed	Mistral	8192
voyage-code-2	Voyage	16000
voyage-code-3	Voyage	32000
gemma2	Ollama	8192
snowflake-arctic-embed	Ollama	8192

Note: For code indexing, models with larger context (8192+ tokens) are recommended to avoid truncation of large functions/classes.

---

Available Tools Reference

Management

Tool	Description
add_folder	Add a folder to be indexed
remove_folder	Remove a folder from indexing
list_folders	List all indexed folders
clear_index	Clear/rebuild index
reindex_folder	Reindex a folder (clears and rebuilds)
pause_indexing	Pause indexing for a folder
resume_indexing	Resume paused indexing
reenrich_folder	Re-enrich folder with LSP (for folders indexed without LSP)

Search

Tool	Description
search	Perform semantic code search

Search Parameters:

• query (required): Natural language search query
• folderPath: Filter to specific folder
• fileTypes: Filter by extensions (e.g., [".ts", ".js"])
• minScore: Minimum similarity (0-1)
• maxResults: Max results to return

Status

Tool	Description
get_status	Get indexing status
get_errors	Get error reports
configure	Update server configuration
get_instructions	Get AI assistant instructions (for CLAUDE.md, etc.)

CLI Usage Examples

# From Claude CLI, you can use these tools directly:
# "search for authentication middleware"
# "add folder /path/to/project"
# "pause indexing for project-name"
# "reindex the api folder"

---

Supported Languages

JavaScript, TypeScript, TSX, JSX, Python, Rust, Go, C, C++, C#, Java, Ruby, PHP, Swift, Kotlin, Scala, Elixir, Erlang, Haskell, OCaml, Lua, Perl, R, Julia, Dart, Vue, Svelte, HTML, CSS, SCSS, SQL, GraphQL, Markdown, JSON, YAML, TOML, XML, Bash, PowerShell, Dockerfile, Terraform, Solidity, Zig, Nim, and more.

---

Ignoring Files

The server respects .gitignore files in indexed folders. You can also create a .cs-mcp-ignore file with additional patterns to exclude.

Always ignored directories:

• node_modules, .git, dist, build, .next, .cache, coverage, .venv, vendor, target, etc.

---

Data Storage

All server data is stored in a platform-appropriate application data folder:

Platform	Location
Windows	%LOCALAPPDATA%\code-search
macOS	~/Library/Application Support/code-search
Linux	~/.local/share/code-search (or $XDG_DATA_HOME/code-search)

Folder contents:

code-search/
├── config.json           # Server configuration
├── folders.json          # Indexed folder registry
├── embedding-cache.db    # SQLite embedding cache
├── lsp-cache.db          # LSP enrichment cache
├── queries/              # Query logs (daily JSONL files)
│   └── queries-YYYY-MM-DD.jsonl
└── cache/                # Per-folder file hash cache
    └── {folder-id}.json

Notes:

• Query logs are automatically cleaned up after 1 day
• Both stdio (Claude CLI) and SSE (Admin UI) servers share the same data
• Embedding cache persists across restarts to avoid redundant API calls

---

Advanced Features

Embedding Cache

The server caches computed embeddings in SQLite to avoid redundant API calls. When re-indexing or updating files:

• Unchanged code chunks reuse cached embeddings
• Only new/modified code requires embedding API calls
• Cache is keyed by content hash and model ID

The cache auto-warms from existing Qdrant data on startup.

LSP Enrichment (Optional)

Enable Language Server Protocol integration to enrich code chunks with type information before embedding:

{
  "lspEnabled": true,
  "lspTimeout": 5000,
  "lspMaxConcurrentRequests": 5,
  "lspUseOmniSharp": false
}

Benefits:

• Type signatures improve semantic matching
• Documentation/JSDoc included in embeddings
• Better results for type-related queries

Prerequisites:

• TypeScript/JavaScript: npm install -g typescript-language-server typescript
• C#: dotnet tool install --global csharp-ls (or set lspUseOmniSharp: true for OmniSharp)

Hierarchical Context

Code chunks automatically include parent context from the AST:

• Class name for methods
• Module name for functions
• Namespace for nested types
• Parent function for nested functions

This helps queries like "authentication method in UserService" match more accurately.

Git Worktree Support

The server automatically detects git worktrees and optimizes indexing:

How it works:

1. When you add a worktree folder, the server detects it's a worktree
2. The base repository is automatically discovered and indexed (if not already)
3. Both repos are indexed, but the embedding cache deduplicates shared code
4. Admin UI shows worktree relationships with "worktree" and "base repo" badges

Benefits:

• No duplicate embedding API calls for shared code (same content = cached embedding)
• Clear visualization of repo relationships in admin UI
• Automatic cleanup detection for deleted worktrees

Important: Keep your base repository checked out to the main/master branch for best results. The index reflects whatever is on disk, not a specific git branch.

Orphaned folder cleanup: If you delete a worktree from disk, the server marks it as "orphaned" on next startup. Use the Cleanup button in the admin UI to remove orphaned indexes.

Reindexing After Upgrade

When upgrading to a version with new enrichment features (LSP, hierarchical context), you need to reindex existing folders to take advantage of the improvements:

1. Open Admin UI at http://localhost:3100/admin
2. Go to the Folders tab
3. Click Reindex on each folder

This clears the existing index and re-indexes all files with the new enrichment features.

---

Future Improvements

See FUTURE-IMPROVEMENTS.md for planned enhancements including:

• LLM re-ranking for improved relevance
• Multi-level code summaries
• Graph-based relationship tracking
• Hybrid BM25 + vector search

---

Troubleshooting

Claude CLI: "Failed to connect" Error

Symptom: claude mcp list shows code-search: ... - ✗ Failed to connect

Solutions:

1. Verify the server can start manually:

   node /path/to/code-search-mcp/dist/index.js
   

   Should output: [CodeSearch] Server running on stdio

2. Check if Qdrant is running:

   curl http://localhost:6333/collections
   

   If not running, start it: docker run -p 6333:6333 qdrant/qdrant

3. Check if Ollama is running (if using Ollama embeddings):

   curl http://localhost:11434/api/tags
   

   If not running, start Ollama and ensure the model is pulled: ollama pull nomic-embed-text

4. Wrong config file location:

   • Claude CLI reads .mcp.json from the project root, NOT .claude/mcp.json
   • Use claude mcp add command to ensure correct placement
   • Run claude mcp list to verify registration

5. SSE vs stdio mode conflict:

   • If running the Admin UI in SSE mode (--sse flag), Claude CLI cannot connect
   • Stop the SSE server before using stdio mode with Claude CLI
   • Or run two separate instances (different ports)

6. Rebuild after code changes:

   npm run build
   

Indexing Stuck or Not Progressing

1. Check for errors in server output:

   • Run manually to see logs: node dist/index.js
   • Look for embedding API errors (context length, rate limits)

2. Embedding model context limits:

   • mxbai-embed-large: 512 tokens (small, may truncate code)
   • nomic-embed-text: 8192 tokens (recommended for code)
   • Large code chunks are automatically truncated with a warning

3. Pause and resume indexing:

   • Use Admin UI or MCP tools to pause/resume
   • Check get_errors for specific file failures

LSP Enrichment Not Working

1. Check LSP server is installed:

   # TypeScript/JavaScript
   typescript-language-server --version
   
   # C#
   csharp-ls --version
   

2. Enable LSP in config:

   { "lspEnabled": true }
   

3. Check LSP status in Admin UI: Services tab shows LSP server status

Search Returns Poor Results

1. Reindex with LSP enabled for better type information
2. Use descriptive queries: "function that validates user email" instead of "validate"
3. Check folder is fully indexed: Use get_status or Admin UI
4. Try different embedding model: Some models work better for code

Windows Path Issues

1. Use forward slashes in JSON config:

   "args": ["C:/repos/code-search/dist/index.js"]
   

   Not backslashes: "C:\\repos\\..." (can cause escaping issues)

2. Avoid spaces in paths or ensure proper quoting

Port Already in Use (SSE Mode)

# Find what's using port 3100
netstat -ano | findstr :3100  # Windows
lsof -i :3100                 # macOS/Linux

# Use a different port
node dist/index.js --sse --port 3101

Qdrant Connection Refused

1. Check Docker is running:

   docker ps | grep qdrant
   

2. Restart Qdrant container:

   docker restart $(docker ps -q --filter ancestor=qdrant/qdrant)
   # Or start fresh:
   docker run -d -p 6333:6333 -v qdrant_storage:/qdrant/storage qdrant/qdrant
   

3. Check firewall/antivirus isn't blocking port 6333

Memory Issues with Large Repositories

1. Increase Node.js memory limit:

   node --max-old-space-size=4096 dist/index.js
   

2. Index folders incrementally instead of all at once

3. Use .cs-mcp-ignore to exclude large generated files or vendor directories

Embedding API Rate Limits

1. Use Ollama for local embeddings (no rate limits)
2. Pause and resume indexing to spread out API calls
3. Check provider dashboard for rate limit details
4. The embedding cache prevents redundant calls on re-indexing

---

License

MIT License - see LICENSE file.

Third-Party Licenses

See THIRD-PARTY-LICENSES.md for dependency licenses.