claude-context — Code Search MCP Server for Claude Code

Introduction

claude-context is an MCP (Model Context Protocol) server built by Zilliz that indexes your codebase into vector embeddings and exposes semantic search as a tool for Claude Code. Instead of relying on grep or file-by-file scanning, Claude can query your entire codebase by meaning, finding relevant code even when exact keywords differ.

What claude-context Does

• Indexes source code files into vector embeddings for semantic search
• Exposes search capabilities to Claude Code via the MCP protocol
• Returns ranked code snippets with file paths and line numbers for each query
• Watches for file changes and incrementally updates the index
• Supports filtering by file type, directory, and language

Architecture Overview

claude-context runs as a local MCP server that connects to Claude Code. On startup, it walks the project directory, chunks source files into meaningful segments, and embeds them using a local or remote embedding model. The embeddings are stored in a Milvus Lite vector database. When Claude issues a search query, the server embeds the query, performs approximate nearest neighbor search, and returns the top-k most relevant code snippets with their locations.

Self-Hosting & Configuration

• Run via npx for zero-install setup or install globally with npm
• Point it at any local project directory to begin indexing
• Configure embedding model choice via environment variables
• Set chunk size and overlap parameters for different codebase scales
• Add to your Claude Code MCP configuration for automatic startup

Key Features

• Semantic code search that understands intent, not just keywords
• Built on Milvus Lite for efficient local vector storage and retrieval
• Incremental indexing that updates only changed files
• Language-aware chunking that respects function and class boundaries
• Low latency retrieval suitable for interactive coding sessions

Comparison with Similar Tools

• Codebase Memory MCP — Builds a knowledge graph of code relationships, while claude-context focuses on vector-based semantic search
• Sourcegraph — Full-featured code intelligence platform, while claude-context is a lightweight local-first MCP server
• grep/ripgrep — Exact string matching only, while claude-context finds semantically similar code regardless of naming
• GitHub Copilot — Integrated into editors with its own retrieval, while claude-context is an open MCP server for Claude Code

FAQ

Q: How large a codebase can it handle? A: claude-context handles repositories with hundreds of thousands of files efficiently using Milvus Lite's optimized indexing.

Q: Does it send my code to external servers? A: When using a local embedding model, all processing happens on your machine. Remote embedding APIs send code chunks to the provider.

Q: How do I add it to Claude Code? A: Add the server configuration to your Claude Code MCP settings file, and it starts automatically when you open a project.

Q: Does it support monorepos? A: Yes, you can configure it to index specific subdirectories or the entire monorepo, with filtering by path prefix.

Sources

• https://github.com/zilliztech/claude-context
• https://milvus.io/