Clients want to use cheaper models, and we wanted to see how can MCPs be designed such that the cheapest of the models can use them the same way a SOTA would. Read our latest blog:
https://t.co/yjmmKO0FOI
Did the harness even call your MCP tool? How many times does it really call the tool vs how many times does it answer from memory? How can you change your tool description to make sure it always calls your tool? M3 helps you answer all these questions.
When your MCP is guarded with tests written using M3, you don’t have to worry about manually testing it for any harnesses. Whether you’re testing with an entirely new one or testing with the latest version of an existing one, its all seamless.
We know that testing MCPs with real agents and harnesses is a pain, and M3 helps you make that process seamless. Just give your agent the M3 skill, and see it write tests using M3, tests that work across Claude code and Codex, any version.