I constructed this after gazing my Claude Code setup burn ~60,000 tokens simply loading instrument definitions from 4 MCP servers—ahead of I even typed a advised.
The issue: MCP provides brokers get admission to to loads of equipment, however each and every instrument description eats context. Redis revealed knowledge appearing the similar factor—167 equipment, 42% variety accuracy, 60K tokens of overhead consistent with consultation. In manufacturing, it may hit 150K+ tokens. The agent spends extra time deciding what to make use of than in reality fixing your drawback.
Present answers fell into two buckets:
– Handbook whitelists (mcpwrapped, MCP Funnel): You need to know upfront which equipment to cover. With 100+ equipment throughout more than one servers, that is a part-time task.
– Business platforms (Stacklok, Redis): Superb accuracy (Stacklok hits 94%), however closed-source or require Redis infrastructure.
I sought after one thing that:
1. Works out of the field with my current claude_desktop_config.json
2. Figures out which equipment I want in keeping with what I am in reality seeking to do
3. Runs 100% in the community, no API keys, no telemetry
4. Is open-source and useless easy to give a contribution to
So I constructed `shutup-mcp` in about 200 strains of Python. It aggregates equipment from your whole MCP servers, builds a neighborhood embedding index (all-MiniLM-L6-v2, ~80MB), and filters equipment via intent ahead of the agent sees them. Token relief in my trying out is ~98%—very similar to what Redis and Atlassian reported with related approaches.
It is v0.1.0, so there is lots left to construct (dynamic re-filtering consistent with message is subsequent). I would love comments from someone else hitting MCP instrument sprawl—what is your present workaround? Regex filters? Specialised brokers? Simply consuming the token value?
Thank you for checking it out. Glad to respond to any questions within the feedback.
