Install the M3 agent skill

The testing-with-m3 skill tells a coding agent what M3 is, how to write and run a first direct test, and which M3 documentation page to read next. Its references/ directory is a copy of the M3 documentation for one release, so the agent reads documentation that matches the M3 you installed.

Automatic installation

m3 init and m3 setup install the skill into the project when it is missing, and update it when it was installed for a different M3 release. They run this command from the project root, with VERSION replaced by the output of m3 --version:

npx --yes [email protected] add sineframe/m3#vVERSION --skill testing-with-m3 --agent universal claude-code -y

The command needs Node.js and Git. It installs the skill only at .agents/skills/testing-with-m3/, creates a .claude/skills/testing-with-m3 link, and records the source and release in skills-lock.json. The skills CLI version is pinned to 1.7.0, and telemetry is disabled for this command. Commit those files if every contributor should get the skill.

M3 does not run the command when:

• you pass --no-skill to m3 init or m3 setup;

• any of CI, GITHUB_ACTIONS, GITLAB_CI, BUILDKITE, CIRCLECI, JENKINS_URL, TF_BUILD, TEAMCITY_VERSION, BITBUCKET_BUILD_NUMBER, or CODEBUILD_BUILD_ID is set to a non-empty value;

• you run m3 setup in a directory without m3.toml;

• M3 is installed for development (editable). Installation is skipped, and M3 prints Agent skill: skipped for a development install of M3. To install it, run: followed by:

  npx --yes [email protected] add sineframe/m3#vVERSION --skill testing-with-m3 --agent universal claude-code -y

• a testing-with-m3 skill that M3 did not install already exists in the project or in your home directory. M3 prints the command to replace it.

If npx is missing or the command fails, M3 prints the command and finishes setup normally.

Install it yourself

Run the command above with your M3 version. To install it globally, use:

npx --yes [email protected] add sineframe/m3#vVERSION --skill testing-with-m3 --agent universal claude-code -g

M3 does not update a global installation, so rerun the command after upgrading M3.

Read the skill source.

Prepare a directory submission

The preparing-directory-submissions skill is for builders submitting an MCP server to the Claude connector directory, the OpenAI plugin directory (ChatGPT and Codex) or the Muse connector directory. Ask your agent to get the server ready for a directory, or paste the rejection feedback you received. The agent then:

• runs m3 lint --directory and carries its findings over unchanged;
• does the checks that need judgment, such as steering language in descriptions, hints that don't match what the code does, catch-all tools, data requests, unsupported use cases, privacy, listing text, connectivity and Muse tool classes, and backs each one with a quote, a file:line reference or command output;
• drafts the review test cases each directory asks for and writes an M3 direct test for each positive case. It runs only the tests for read-only tools unless you confirm a test account;
• writes directory-readiness-report.md, with a table per directory that quotes the policy line behind every failure and proposes fixes as diffs.

The agent never edits your server without asking, never submits anything, and never asks for credentials. The report says what was checked and what was found. It can't tell you whether a directory will approve the server.

m3 init and m3 setup don't install this skill. Install it once per project with the same command as testing-with-m3, with VERSION replaced by the output of m3 --version:

npx --yes [email protected] add sineframe/m3#vVERSION --skill preparing-directory-submissions --agent universal claude-code -y

It installs at .agents/skills/preparing-directory-submissions/, next to testing-with-m3, and adds its own entry to skills-lock.json. It doesn't change testing-with-m3, and m3 setup keeps managing testing-with-m3 as before. Use -g in place of -y to install it globally. Rerun the command after upgrading M3; the skill's policy excerpts are dated, and newer releases carry newer ones.

The skill is available from the first M3 release that includes it, so VERSION must be that release or later. If the m3 command the agent finds has no m3 lint --directory, for example an older installation earlier on PATH, the agent still runs the judgment checks, marks the lint rows "not checked" and tells you how to upgrade.

Read the skill source.