AI product
SineFrame M3 is a pytest framework for testing MCP servers with direct server checks and real agent harnesses, including Claude Code, Codex, OpenCode, and Pi. It runs ordinary pytest files against the harness binaries, records every MCP tool call, and lets tests assert that an agent used the expected tool with the expected arguments or that a server returned the required schema, response, error, or state. M3 can compare the same suite across agents, model or runtime versions, and custom agents that expose the Agent Client Protocol (ACP); agents without ACP require a wrapper process.
The CLI runs tests locally with `m3 test` and in CI with `m3 ci test`, failing a test when an MCP server is skipped, called incorrectly, or returns an invalid result. It also fails the CI gate when no tests execute. Run history is stored in a local `.m3/executions.sqlite` database, reports are written to `.m3/reports`, and an optional local viewer can be opened with `m3 test --ui`. The project provides an agent skill for generating tests and can be installed manually with `m3 init`, `m3 setup`, and `m3 doctor`.
1 use taken from transcripts — each links to the moment in the video.
A pytest-based framework that tests MCP servers against real agents like Claude Code, Codex, and OpenCode, recording every MCP tool call and failing tests when the server is skipped or called wrong, with CI gating and local-first open source.
1 in the library.