Summary

DeepContext tackles the challenge of providing coding agents with accurate context for large codebases, where traditional grep-based search fails. It offers an MCP server that integrates symbol-aware semantic search, enabling agents like Claude Code and Codex CLI to precisely understand vast code repositories. This leads to improved agent performance, reduced token consumption, and more reliable code completion and analysis for complex projects.

How It Works

DeepContext uses an MCP server architecture with Tree-sitter for AST-based parsing of Python and TypeScript, creating semantic code chunks. Search is a three-stage hybrid process: vector similarity (Jina embeddings) combined with BM25 full-text search, followed by Jina reranker-v2. Incremental indexing, driven by file modification times and SHA-256 hashes, efficiently updates by reprocessing only changed files. Content filtering excludes non-source code artifacts.

Quick Start & Requirements

Installation for MCP clients like Claude Code involves adding the DeepContext MCP server with an API key. Codex CLI integration requires configuration in ~/.codex/config.toml. For local self-hosting, clone the repo, install Node.js dependencies (npm install), build (npm run build), and integrate the standalone MCP script using Turbopuffer and Jina API keys.

• Prerequisites: Node.js 20+, Wildcard API Key (hosted), Turbopuffer API Key (self-hosted), Jina AI API Key (self-hosted).
• Links: Demo: https://github.com/user-attachments/assets/9a2d418f-497b-42b9-bbb2-f875ef0007b4.

Highlighted Details

• Symbol-aware semantic search for precise code context.
• AST-based parsing via Tree-sitter.
• Hybrid search (vector + BM25) with advanced reranking.
• Incremental indexing for efficient updates.
• Content filtering for source code focus.

Maintenance & Community

DeepContext is an early-stage startup project. External contributions are not currently accepted to facilitate rapid iteration. Users interested in future developments should star the repository. No community channels are listed.

Licensing & Compatibility

Licensed under the Apache License 2.0, this project is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Currently supports only TypeScript and Python. Self-hosting requires code modifications to integrate directly with external vector storage and embedding providers, as the default implementation uses the Wildcard API backend. The project is in an early stage and not accepting external contributions.