@turbo-proxy/mcp

MCP Memory Server for Claude Code — provides persistent memory, code intelligence, and project scanning tools via the Model Context Protocol.

Quick Start

# Install globally
npm install -g @turbo-proxy/mcp

# Setup in your project (creates .mcp.json + .claude/settings.json)
turbo-proxy-mcp setup --api-key YOUR_PROXY_KEY --backend-url http://localhost:2300

# Restart Claude Code — memory tools are now available

Requirements

• AIProxy backend running with memory enabled (memory.enabled: true in config.yaml)
• API key configured in proxy's api-keys section

What It Does

When you start a Claude Code session in a configured project:

1. SessionStart hook loads previous memories from the backend
2. 13 MCP tools are available for Claude to use during the session
3. Memories persist across sessions via the proxy backend (MongoDB)

Tools

Memory (6 tools)

| Tool | Description | |------|-------------| | memory_store | Save a memory (fact, preference, decision, pattern, convention) | | memory_recall | Search memories by natural language query | | memory_context | Load full project context for session initialization | | memory_search | Advanced search with filters and pagination | | memory_list | List memories with sorting | | memory_forget | Delete a memory |

Code Intelligence (3 tools)

| Tool | Description | |------|-------------| | code_index | Build/update code graph (incremental, SHA-256 diff) | | code_blast_radius | Trace affected files when code changes | | code_search | Search code graph symbols |

Scanners (5 tools)

| Tool | Description | |------|-------------| | scan_project | File tree, stack detection, monorepo detection | | scan_git | Commits, branches, contributors, hotspots | | scan_changes | Git diff analysis, pattern detection | | scan_todos | TODO/FIXME/HACK/XXX/NOTE grep | | scan_errors | Typecheck + lint wrapper (timeout-bound) |

Configuration

Environment Variables

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | MCP_BACKEND_URL | Yes | http://localhost:2300 | Proxy backend URL | | MCP_API_KEY | Yes | — | Proxy API key | | MCP_LOG_LEVEL | No | info | Log level (debug, info, warn, error) |

Manual Setup

If you prefer manual configuration over turbo-proxy-mcp setup:

.mcp.json (project root):

{
  "mcpServers": {
    "memory": {
      "command": "cmd",
      "args": ["/c", "npx", "@turbo-proxy/mcp"],
      "env": {
        "MCP_BACKEND_URL": "http://localhost:2300",
        "MCP_API_KEY": "your-key"
      }
    }
  }
}

On macOS/Linux, use "command": "npx" and "args": ["@turbo-proxy/mcp"] instead.

Architecture

Claude Code ←stdio→ MCP Server ←HTTP→ Proxy Backend (port 2300) ←→ MongoDB
                       ↓
                  Local SQLite
                  (cache + codegraph)

• Memory tools → HTTP calls to proxy backend REST API
• Code intelligence → Local only (web-tree-sitter WASM + sql.js)
• Scanners → Local only (filesystem + git commands)
• Offline mode → LocalCache + SyncQueue with exponential backoff

Zero Native Dependencies

All heavy lifting uses WASM:

• sql.js — SQLite via WASM (cache, codegraph, audit log)
• web-tree-sitter — AST parsing via WASM (code indexing)

No native build tools required. Works on Windows, macOS, Linux.

License

Proprietary — part of the AIProxy project.