Ended up throwing Multipass VMs at it with 8 cores. Not elegant, not clever caching, just brute force compute. But it works now and I can actually test things locally before pushing.

Sometimes the solution is just "use more CPU" and that's fine.

Yeah, I asked the question on Reddit and got something useful www.reddit.com/r/mcp/commen...

The outcome: I know exactly which AIs can successfully use my MCP server and which ones can't. If none of them can, I know my tool descriptions or schema need work before any real user hits this wall.

I really want this to exist.

Then it checks: Does each model understand what the user actually wants? Does it understand my MCP's capabilities? Does it prepare the data correctly? Does it call my tools in the right sequence with the right parameters?

Here's what I want:

I give it real user input ("find me all TypeScript files modified this week") and my MCP implementation. The test runs this through multiple AI models—Claude, GPT, Gemini, whatever I want to support.

Ex-Amazon SDE here. I love building side projects, but managing deployments manually is a headache.

With AI writing more code faster, I needed a better solution. If Claude can build my service, why can't it deploy it too?

⭐ github.com/ottershipper...
Follow for updates →