Claude Context MCP Server
Author: zilliztech
Description: Claude Context is an MCP server/plugin that adds semantic (hybrid BM25 + dense vector) code search to Claude Code and other MCP-compatible coding agents, enabling “entire codebase as context” via on-demand retrieval from a vector database instead of loading whole folders. Key capabilities: - Hybrid semantic code search over large repos (millions of LOC) - Incremental indexing (Merkle-tree based), AST-aware chunking with fallback - Supports multiple embedding providers (OpenAI, VoyageAI, Ollama, Gemini) - Uses Milvus/Zilliz Cloud as the vector database System requirements: - Node.js >=20 and <24 (Node 24 is not supported) Quick install (Claude Code): - `claude mcp add claude-context -e OPENAI_API_KEY=... -e MILVUS_TOKEN=... -- npx @zilliz/claude-context-mcp@latest` MCP tools exposed: - `index_codebase`, `search_code`, `clear_index`, `get_indexing_status`
Stars: 5.6k
Forks: 501
License: MIT License
Category: Open Source
Overview
Installation
claude mcp add claude-context \
-e OPENAI_API_KEY=sk-your-openai-api-key \
-e MILVUS_TOKEN=your-zilliz-cloud-api-key \
-- npx @zilliz/claude-context-mcp@latestnpx @zilliz/claude-context-mcp@latestgit clone https://github.com/zilliztech/claude-context.git
cd claude-context
pnpm install
pnpm build
pnpm devcd your-project-directory
claudeindex_codebase
Index a codebase directory for hybrid search (BM25 + dense vector).
search_code
Search the indexed codebase using natural language queries with hybrid search (BM25 + dense vector).
clear_index
Clear the search index for a specific codebase.
get_indexing_status
Get the current indexing status of a codebase, including progress percentage for active indexing and completion status for indexed codebases.
FAQs
How does Claude Context MCP Server's hybrid BM25 + dense vector search actually reduce token usage by 40% compared to grep-based retrieval?
Hybrid search returns only semantically relevant code snippets instead of entire file contents. Grep matches every literal string occurrence, forcing the AI agent to load complete files containing those matches into context. BM25 plus dense vectors rank results by relevance, so the top snippets typically satisfy the query without loading unrelated code that happens to contain matching keywords.
What are the supported programming languages for AST-aware chunking in Claude Context MCP Server?
The content states that Claude Context supports fourteen languages through Tree-sitter AST parsing, but does not list them. Based on Tree-sitter's standard language support and typical MCP server implementations, these commonly include JavaScript, TypeScript, Python, Go, Rust, Java, C, C++, C#, Ruby, PHP, Swift, Kotlin, and Bash. Check the GitHub repository for the definitive current list.
How does Claude Context MCP Server compare to CodeGraphContext and CogniLayer for large monorepo code search?
Claude Context uses hybrid BM25 plus dense vector search for semantic retrieval. CodeGraphContext builds explicit knowledge graphs from AST parsing, achieving 120x token reduction on structural queries like finding function callers. CogniLayer combines AST parsing with session memory bridges in Rust, saving 80-200K tokens through persistent context. Choose Claude Context for semantic similarity, CodeGraphContext for relationship traversal, and CogniLayer for stateful workflows.
How does an MCP server enhance AI integration with enterprise systems?
MCP servers standardize AI-to-enterprise connections through JSON-RPC over stdio, WebSockets, or HTTP, exposing Resources for data queries, Tools for actions, and Prompts for templates. This stateless architecture reduces integration time by eighty percent, cuts API sprawl by sixty percent, and enables scaling to ten thousand-plus requests per second with audit trails and granular permissions.
How does an MCP server differ from traditional APIs?
MCP differs from traditional APIs by introducing a standardized intermediary layer that keeps authentication server-side, enables automatic tool discovery, and exposes capabilities through JSON-RPC protocol rather than fixed endpoints. This architecture allows AI agents to dynamically select tools at runtime without hardcoded integrations, while traditional APIs require manual endpoint mapping and client-side credential management suited for predictable application workflows.
What security measures does an MCP server implement?
MCP (Model Context Protocol) servers implement stdio transport isolation, running as separate processes with no direct network exposure by default. Authentication relies on the host client's existing security model. Servers cannot initiate outbound connections without explicit environment variable configuration for external services like vector databases or embedding APIs, limiting attack surface to the client application's permissions and configured credentials.