Projects / MCP Playground
MCP Playground
This is a v1 simulation that mirrors how the actual Runloop-MCP code behaves: stdio MCP server, typed tool registration, and predictable tool-response envelopes.
Prompt: Create a devbox, run test commands, and return a status summary for the run.
Press Run to start the trace.
Origins
I built this project when Anthropic first released MCP and it was immediately clear that agent tooling needed a shared interface. Before MCP, every integration felt custom and brittle. The same model behavior had to be re-wired repeatedly for each platform, which made iteration slow and debugging harder than it needed to be.
MCP changed that by giving us a consistent way to expose tools: named methods, structured parameters, and predictable responses. In plain terms, it gave the model a cleaner contract with the outside world. That consistency is what made this Runloop integration practical: instead of inventing a one-off bridge, I could build one MCP server surface and map tool groups to real Runloop capabilities.
In Runloop-MCP, that meant wiring common workflows like devbox lifecycle operations, blueprint management, repository inspection, scenario orchestration, and benchmark execution into one protocol boundary. The server in src/index.ts keeps the entrypoint simple, while tool modules handle domain behavior. Schema definitions from runLoopAPI.ts are used directly for validation so tool calls stay aligned with API contracts.
The result was a working bridge between model intent and infrastructure actions: creating environments, running commands, gathering logs, and driving evaluation flows through a single MCP interface.
Since then, Runloop has launched an MCP Hub (docs.runloop.ai/docs/devboxes/mcp-hub). I can’t precisely map which parts trace directly to my implementation, but this project was my early foundation for integrating MCP into Runloop workflows and proving the pattern end-to-end.