I built this after watching my Claude Code setup burn ~60,000 tokens just loading tool definitions from four MCP servers—before I even typed a prompt. The problem: MCP gives agents access to hundreds of tools, but every tool description eats context. Redis published data showing the same thing—167 tools, 42% selection accuracy, 60K tokens of overhead per session. In production, it can hit 150K+ tokens. The agent spends more time deciding what to use than actually solving your problem. Existing solutions fell into two buckets: - Manual whitelists (mcpwrapped, MCP Funnel): You have to know in advance which tools to hide. With 100+ tools across multiple servers, that's a part-time job. - Commercial platforms (Stacklok, Redis): Excellent accuracy (Stacklok hits 94%), but closed-source or require Redis infrastructure. I wanted something that: 1. Works out of the box with my existing claude_desktop_config.json 2. Figures out which tools I need based on what I'm actually trying to do 3. Runs 100% locally, no API keys, no telemetry 4. Is open-source and dead simple to contribute to So I built `shutup-mcp` in about 200 lines of Python. It aggregates tools from all your MCP servers, builds a local embedding index (all-MiniLM-L6-v2, ~80MB), and filters tools by intent before the agent sees them. Token reduction in my testing is ~98%—similar to what Redis and Atlassian reported with comparable approaches. It's v0.1.0, so there's plenty left to build (dynamic re-filtering per message is next). I'd love feedback from anyone else hitting MCP tool sprawl—what's your current workaround? Regex filters? Specialized agents? Just eating the token cost? Thanks for checking it out. Happy to answer any questions in the comments.

shutup-mcp

Zero-config MCP proxy that hides 99% of tools

Zero-config MCP proxy that hides 99% of tools