Summary
A second opinion on tool and model choice, cheap enough to ask every turn.
Hardcoding one model and one scraper at build time is the thing that quietly costs money as the task mix changes. ToolRoute answers per task and takes your reported outcomes back, so the answer drifts toward what works on your own workload rather than a public benchmark.
What it is
A hosted routing layer for agents. You describe a task, it names the MCP server and the model to use and returns a confidence score; you still make the call yourself.
What you get
- A recommended MCP server plus a recommended model for one task, with alternatives and a fallback
- Model-only routing when the task needs no external tool
- An ordered chain for compound tasks that touch several tools
- Outcome reporting, so later recommendations reflect what actually worked on your traffic
- A catalog search over the MCP servers it knows about
- A health endpoint and public aggregate metrics
Requirements
Nothing to install for the MCP path — point a client at the hosted endpoint. The SDK is published as @toolroute/sdk; the registry entry io.github.grossiweb/toolroute installs @toolroute/sdk over stdio. Self-hosting needs Supabase credentials.
Setup effort
One command — npx -y @toolroute/sdk@0.2.2 toolroute-mcp