Model Context Protocol
One tool returns { user: { id: number } }. The next one wants { userId: string }. Nothing tells you until it breaks in production. toolgraph checks every connection against the tools' real JSON Schemas the moment you draw it.
createUser
users server
Output
sendEmail
mail server
Input
That connection would not type-check
Field userId expects string, but createUser provides number.
Every connection is validated against both tools' real schemas before the edge exists. Incompatible ones are refused with the field, the expected type and the actual type — not a generic error.
Test the whole graph against the live servers and watch each step stream back as it finishes. No mock layer between you and what will actually happen.
Generate standalone TypeScript with real interfaces and zod validators, or Python with Pydantic models. Zero toolgraph dependency — delete your account and the code keeps working.
The whole thing is open source and self-hostable. Nothing here is a proprietary runtime you would have to migrate off later.
One
Point toolgraph at an MCP server over streamable HTTP or SSE. It reads the tools and their schemas directly from the protocol.
Two
Drop tools on the canvas and connect field to field. Anything that would not type-check is refused, with the mismatch named inline.
Three
Test against the real servers, then export the graph as TypeScript or Python and run it wherever you like.
