You want to test the full request flow between client and server: connection, tool discovery, tool selection, calls, responses, and whether the user’s GOAL was actually achieved.
What's hard: an external agent sits between your user and your MCP server, in a client you don't control (ChatGPT, claude etc.). Your tests need to cover how that client and agent find and uses your tools.
Recommend checking for :
- (deterministic assertions + non-deterministic judge checks) at essentially every stage of the request flow from client->server (User Value Chain)
- across many clients where YOUR target users are at (unfort. these clients change behavior every other day) and across harness + models
- our default eval assertions cover things like input schemas, valid arguments, repeated calls, errors, latency, and response size; pair those deterministic checks with non-deterministic evals for tool choice, response interpretation, and task completion.
We also think giving people completely configurable clients is extremely valuable, we keep these client "templates" up to date with the latest capabilities- but folks can certainly change up protocol/extension capabilities beyond that e.g. if Copilot is planning to support more apps capabilities in the coming weeks, use the Copilot template and toggle on more MCP apps endpoints to see if you're able to make use of them
Prathmesh, CEO of MCPJam here.Users now start in ChatGPT, Claude, Cursor, and other AI clients. They reach your product through your MCP server.That means your users often aren’t in your product anymore. You can’t see what they prompted for, how the agent interpreted it, or whether your server helped them get the result they wanted.I saw this firsthand leading MCP technical strategy at Asana,...