Mcp Spine

What is Mcp Spine?

Mcp Spine is a local-first proxy that sits between an MCP client (such as Claude Desktop) and your MCP servers. It adds security, routing, token control, and compliance — handling what goes in, what comes out, and what gets logged — all through a single configuration file.

How to use Mcp Spine?

Install with pip install mcp-spine (add [ml] for semantic routing). Run mcp-spine init for an interactive setup wizard, then start the proxy with mcp-spine serve --config spine.toml. Replace individual MCP server entries in your client configuration with a single Spine entry pointing to the proxy.

Key features of Mcp Spine

• Security proxy with rate limiting, secret scrubbing, and path jails
• Semantic router using local embeddings (no API calls)
• Schema minifier achieving 61% token savings at default level
• Human-in-the-loop confirmation for destructive tools
• Prompt injection detection in tool responses
• Multi-user audit trail with session tagging
• Web dashboard for live monitoring and analytics

Use cases of Mcp Spine

• Centralize security policies across multiple MCP servers
• Reduce token consumption by minifying tool schemas and routing only relevant tools
• Enforce daily token budgets per user or server
• Audit all tool calls with session tracking for shared deployments
• Prevent prompt injection and command injection in agent workflows

FAQ from Mcp Spine

How is Mcp Spine different from running MCP servers directly?

It acts as a middleware proxy that adds security, routing, token control, and audit logging between the client and servers — capabilities not available when connecting tools directly.

What are the runtime requirements?

Python 3.8+ is required. The base install uses only standard library dependencies; optional ML dependencies (pip install mcp-spine[ml]) enable semantic routing.

Where does audit data live?

All audit events, token budgets, and session data are stored in a local SQLite database (default spine_audit.db). No data leaves your machine.

What transport protocols are supported?

stdio (local subprocess servers), SSE (legacy remote servers), and Streamable HTTP (MCP 2025-03-26 spec).

Does semantic routing require external API calls?

No. It uses the local embedding model all-MiniLM-L6-v2 with ChromaDB for tool indexing — all processing stays on your machine.