Claude Context MCP Server: Persistent Markdown Memory

Overview

Claude Context MCP Server implements a lightweight Model Context Protocol (MCP) plugin that gives Claude Code and other MCP-compatible AI coding assistants persistent, markdown-first memory backed by a vector database. Instead of loading entire directories into the model on every request, Claude Context indexes your codebase (and other documentation) as embeddings and performs semantic search to return only the most relevant snippets into Claude’s context. This makes context richer, more precise, and far more cost-effective for large codebases.

The server integrates with memsearch-style workflows to store and retrieve Markdown “memory” items that persist across sessions. It exposes a simple MCP server binary (typically run via npx) that MCP-capable clients (Claude Code, Gemini CLI, Qwen, Codex CLI, Cursor, etc.) can spawn and talk to, enabling session-spanning memory and deeper developer workflows.

Features

• Persistent Markdown-first memory for Claude Code and other MCP clients
• Semantic code search powered by vector embeddings and a Milvus-compatible vector DB
• Cost-efficient context: only relevant snippets are inserted into model prompts
• Easy MCP integration via npx / MCP configuration files (JSON, TOML)
• Support for OpenAI embeddings (or other compatible embedding providers)
• CLI-friendly setup and configurable startup timeout and environment variables

Installation / Configuration

Prerequisites:

• Node.js >= 20.0.0 and < 24.0.0
• A vector database account (Milvus / Zilliz Cloud) and API token
• An embedding provider API key (e.g., OpenAI API key for embeddings)

Quick install (run as MCP server from a client like Claude Code):

# Add Claude Context as an MCP server to Claude Code (example)
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@latest

Environment variables (examples):

• OPENAI_API_KEY — API key used for generating embeddings
• MILVUS_TOKEN — Zilliz Cloud / Milvus token for vector DB access
• MILVUS_ADDRESS — (optional) public endpoint for Milvus if needed

Recommended MCP client configuration examples

Codex CLI (TOML):

[mcp_servers.claude-context]
command = "npx"
args = ["@zilliz/claude-context-mcp@latest"]
env = { "OPENAI_API_KEY" = "sk-your-openai-api-key", "MILVUS_TOKEN" = "your-zilliz-cloud-api-key" }
startup_timeout_ms = 20000

Gemini CLI (JSON):

{
  "mcpServers": {
    "claude-context": {
      "command": "npx",
      "args": ["@zilliz/claude-context-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-api-key",
        "MILVUS_TOKEN": "your-zilliz-cloud-api-key"
      }
    }
  }
}

Qwen (JSON):

{
  "mcpServers": {
    "claude-context": {
      "command": "npx",
      "args": ["@zilliz/claude-context-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-api-key",
        "MILVUS_ADDRESS": "your-zilliz-cloud-public-endpoint",
        "MILVUS_TOKEN": "your-zilliz-cloud-api-key"
      }
    }
  }
}

Available Resources

• GitHub repository: https://github.com/zilliztech/claude-context
• NPM packages:
  • @zilliz/claude-context-core
  • @zilliz/claude-context-mcp
• Docs / DeepWiki: project documentation and guides (see repository README and docs/ folder)
• Memsearch plugin for Claude Code: markdown-first memory ingestion and indexing
• Zilliz Cloud: managed Milvus / vector DB service for storing embeddings

Environment variables reference:

Use Cases

• Large monorepo code search: Index millions of lines of code and let Claude Code retrieve relevant snippets for bug fixes, refactors, or design questions without loading whole files.
• Cross-session developer memory: Save design notes, TODOs, architecture decisions, and pair-programming transcripts as Markdown entries that persist between sessions and are retrieved by semantic relevance.
• Agentic RAG (retrieval-augmented generation): Combine retrieval from the vector DB with Claude’s reasoning for code generation, test creation, or security reviews.
• IDE augmentation: Wire Claude Context into your CLI/IDE extension via MCP so in-editor assistant sessions carry long-term project context.

Concrete example: Bug triage

1. Index repository and engineering docs into the vector DB.
2. In Claude Code, ask “Why does endpoint /payments return 500 on invalid card?” Claude Context semantically finds relevant handlers, logs, and docs and inserts concise snippets into the model prompt.
3. Claude returns a diagnosis with links to the exact files and code sections (no manual file digging).

Tips & Troubleshooting

• Node version: ensure Node >=20 and <24. If you’re on Node 24+, downgrade or use an npx wrapper that sets a compatible runtime.
• Startup timeout: some MCP clients have a short default startup timeout. Use startup_timeout_ms (e.g., 20000) in configuration if you see startup failures.
• Cost control: tune the number of retrieved snippets and maximum token budget in your MCP client to balance context richness and API costs.
• Indexing: use memsearch or the repository’s ingestion utilities to generate Markdown-first memory items (code + context) for better retrieval quality.

For more implementation details, API references, and advanced configuration, see the project repository and the included docs.