Sourcerer MCP 🧙

An MCP server for semantic code search & navigation that helps AI agents work efficiently without burning through costly tokens. Instead of reading entire files, agents can search conceptually and jump directly to the specific functions, classes, and code chunks they need.

Demo

Requirements

• OpenAI API Key: Required for generating embeddings (local embedding support planned)
• Git: Must be a git repository (respects .gitignore files)
• Add .sourcerer/ to .gitignore: This directory stores the embedded vector database

Installation

Go

go install github.com/st3v3nmw/sourcerer-mcp/cmd/sourcerer@latest

Homebrew

brew tap st3v3nmw/tap
brew install st3v3nmw/tap/sourcerer

Configuration

Claude Code

claude mcp add sourcerer -e OPENAI_API_KEY=your-openai-api-key -e SOURCERER_WORKSPACE_ROOT=$(pwd) -- sourcerer

mcp.json

{
  "mcpServers": {
    "sourcerer": {
      "command": "sourcerer",
      "env": {
        "OPENAI_API_KEY": "your-openai-api-key",
        "SOURCERER_WORKSPACE_ROOT": "/path/to/your/project"
      }
    }
  }
}

How it Works

Sourcerer 🧙 builds a semantic search index of your codebase:

1. Code Parsing & Chunking

• Uses Tree-sitter to parse source files into ASTs
• Extracts meaningful chunks (functions, classes, methods, types) with stable IDs
• Each chunk includes source code, location info, and contextual summaries
• Chunk IDs follow the format: file.ext::Type::method

2. File System Integration

• Watches for file changes using fsnotify
• Respects .gitignore files via git check-ignore
• Automatically re-indexes changed files
• Stores metadata to track modification times

3. Vector Database

• Uses chromem-go for persistent vector storage in .sourcerer/db/
• Generates embeddings via OpenAI's API for semantic similarity
• Enables conceptual search rather than just text matching
• Maintains chunks, their embeddings, and metadata

4. MCP Tools

• semantic_search: Find relevant code using semantic search
• get_chunk_code: Retrieve specific chunks by ID
• find_similar_chunks: Find similar chunks
• index_workspace: Manually trigger re-indexing
• get_index_status: Check indexing progress

This approach allows AI agents to find relevant code without reading entire files, dramatically reducing token usage and cognitive load.

Supported Languages

Language support requires writing Tree-sitter queries to identify functions, classes, interfaces, and other code structures for each language.

Supported: Go, JavaScript, Markdown, Python, TypeScript

Planned: C, C++, Java, Ruby, Rust, and others

Contributing

All contributions welcome! See CONTRIBUTING.md.

$ ls @stephenmwangi.com
- gh:st3v3nmw/obsidian-spaced-repetition
- gh:st3v3nmw/lsfr