sineframe.com

mx

mailserver.purelymail.com

ns

ulla.ns.cloudflare.com

sullivan.ns.cloudflare.com

caa

txt

purelymail_ownership_proof=0f3653637e81f0c051412ff60242a0f0c2a8fea8d3a9dcd90f9cdb96f12a19d8f144bfc41c45ef8ca6186c284345969ae2d4312221b36b6da8eda8f2dac4f3a0

google-site-verification=KrspUn5G5IXz92-raRW7eQQNFo31c75ea3sSR22k-iM

v=spf1 include:_spf.purelymail.com ~all

dmarc

v=DMARC1; p=reject; ruf=mailto:dmarc@purelymail.com

User-agent: *
Allow: /

Sitemap: https://m3.sineframe.com/sitemap.xml
Sitemap: https://m3.sineframe.com/docs/sitemap.xml

# SineFrame M3

> M3 tests MCP servers and the agents that use them. Published guidance matches M3 v0.3.1.

Use the agent instructions to add M3 tests to a project. Use the documentation index for CLI, SDK, CI, and evaluation reference material.

## Start here

- [Agent instructions](https://m3.sineframe.com/agents.md): Install M3 and give a coding agent the version-matched M3 testing skill.
- [M3 documentation](https://m3.sineframe.com/docs/llms.txt): Curated CLI, SDK, CI, and evaluation documentation.
- [Source repository](https://github.com/sineframe/m3): Source code, releases, and issue tracker.

## Guides

- [How to test an MCP server](https://m3.sineframe.com/mcp-testing): Direct pytest tests over stdio or HTTP, plus agent tool-selection tests through real Claude Code, Codex, OpenCode, and Pi binaries.
- [Gate an MCP server in CI](https://m3.sineframe.com/mcp-ci): Run the same M3 suite in GitHub Actions with `m3 ci test`; a failed assertion or a run with zero executed checks fails the gate.
- [SineFrame M3 vs MCPJam](https://m3.sineframe.com/compare/mcpjam): A neutral comparison for MCP server testing: authoring, agent coverage, OAuth, CI gates, results, and when MCPJam fits better.

## Blog

- [On a well-built MCP server, Haiku matched Opus](https://m3.sineframe.com/blog/mcp-server-design-small-models.md): Six models (Claude Code 2.1.296, Codex 0.162.1) on the same 13 issue-tracker tasks over a well-built and a flawed MCP server, 1,248 graded runs; where small models break, which flaws compound, and cost per success.
- [Claude Code and Codex break on different MCP features](https://m3.sineframe.com/blog/claude-code-vs-codex-mcp.md): Ten MCP features run against Claude Code 2.1.287 and Codex 0.162.0 at pinned versions, three trials each; five split the agents, plus eight probe-only leads, and how each cell is measured with M3.
- [Your MCP server passed. Did the agent call it?](https://m3.sineframe.com/blog/did-the-agent-call-it.md): A walkthrough with real runs: 64 pytest cases against a shipping-quote MCP server, a one-line regression that fails 17 of them, the same agent tests through Claude Code and Codex, the trace, and the CI gate.

## Optional

- [Full documentation snapshot](https://m3.sineframe.com/docs/llms-full.txt): All published documentation in one Markdown document.