Wispr Flow: Dictation That Works EverywhereStop typing. Start speaking. 4x faster.
Promoted
shutup-mcp
Zero-config MCP proxy that hides 99% of tools
46 followers
Zero-config MCP proxy that hides 99% of tools
46 followers
Zero-config MCP proxy that shows agents only the tools they actually need — 98% token reduction, no API keys. - hjs-spec/shutup-mcp
shutup-mcp
98% token reduction is wild. I've been working with MCP servers and the tool list bloat is a real problem the agent spends half its context just reading tool descriptions. The zero-config approach is smart because the moment you ask devs to configure which tools to expose, you've already lost. How does it decide which tools to surface for a given query?
shutup-mcp
@linoy_bar_gal — thanks for the kind words, and great question.
The proxy extracts intent from each user message, then uses hybrid search to find the most relevant tools. Hybrid means:
Vector embeddings (semantic similarity — understands "read a file" ≈ "open document")
BM25 keyword matching (exact term matching — catches "GitHub" vs "git")
RRF fusion to merge both result sets without score normalization
Industry benchmarks show this pushes tool selection accuracy from ~38% (vector alone) to ~94% (hybrid). Token reduction stays ~98%.
Honest caveat: intent extraction currently happens during tool calls. Pure conversation first steps (no tool invocation yet) still see the full tool set. Closing that gap in a future version.
Curious—in your MCP setups, do you find the agent picking the wrong tool more often, or just burning tokens on descriptions? The failure mode tells you which part of the pipeline needs tuning.
"Not a dashboard, a co-worker" this positioning resonates. As a solo founder, I don't have a data team to investigate why a metric dropped. I just see the drop and panic. The idea of an AI that can connect the dots between a code release and a metric change is exactly what small teams need. How granular does the root cause analysis get?