Modern RAG for your codebase - semantic and regex search via MCP.

Transform your codebase into a searchable knowledge base for AI assistants using semantic search via cAST algorithm and regex search. Integrates with AI assistants via the Model Context Protocol (MCP).

Features

• cAST Algorithm - Research-backed semantic code chunking
• Semantic search - Natural language queries like "find authentication code"
• Regex search - Pattern matching without API keys
• Local-first - Your code stays on your machine
• 22 languages with structured parsing
  • Programming (via Tree-sitter): Python, JavaScript, TypeScript, JSX, TSX, Java, Kotlin, Groovy, C, C++, C#, Go, Rust, Bash, MATLAB, Makefile
  • Configuration (via Tree-sitter): JSON, YAML, TOML, Markdown
  • Text-based (custom parsers): Text files, PDF
• MCP integration - Works with Claude, VS Code, Cursor, Windsurf, Zed, etc

Documentation

Visit ofriw.github.io/chunkhound for complete guides:

• Tutorial
• Configuration Guide
• Architecture Deep Dive

Requirements

• Python 3.10+
• uv package manager
• API key for semantic search (optional - regex search works without any keys)
  • OpenAI | VoyageAI | Local with Ollama

Installation

# Install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install ChunkHound
uv tool install chunkhound

Quick Start

Option 1: With Embeddings (Recommended)

1. Create .chunkhound.json in project root file

{
  "embedding": {
    "provider": "openai",
    "api_key": "your-api-key-here"
  }
}

2. Index your codebase

chunkhound index

Option 2: Without embeddings (regex search only)

chunkhound index --no-embeddings

For configuration, IDE setup, and advanced usage, see the documentation.

Why ChunkHound?

Research Foundation: Built on the cAST (Chunking via Abstract Syntax Trees) algorithm from Carnegie Mellon University, providing:

• 4.3 point gain in Recall@5 on RepoEval retrieval
• 2.67 point gain in Pass@1 on SWE-bench generation
• Structure-aware chunking that preserves code meaning

Local-First Architecture:

• Your code never leaves your machine
• Works offline with Ollama local models
• No per-token charges for large codebases

Universal Language Support:

• Structured parsing for 22 languages (Tree-sitter + custom parsers)
• Same semantic concepts across all programming languages

License

MIT