Use with an AI assistant

mcmc is built for agents as much as for people: every command takes --json, exit codes are meaningful, and each run is recorded. Two pieces make that usable from an assistant: a skill that teaches the conventions, and an MCP server that exposes the commands as tools.

Neither is a separate download. Both ship inside the CLI you already have.

The skill

mcmc skill install             # for you, in ~/.claude/skills
mcmc skill install --project   # for the repo, in ./.claude/skills

It teaches what --help cannot: that a model file must define build_model(data), that the @model macro goes at the top level, how to read a not-converged verdict, and the discipline of checking a model with a prior predictive check and mcmc sbc before trusting it. mcmc skill show prints it without installing.

Installing into the project puts the skill under version control, so everyone who clones the repo gets the same guidance.

The tools

claude mcp add mcmcjs -- mcmc mcp

That registers mcmc mcp, which speaks the Model Context Protocol on stdin and stdout. For a team, commit an .mcp.json instead:

{
  "mcpServers": {
    "mcmcjs": { "command": "mcmc", "args": ["mcp"] }
  }
}

The server offers eight tools:

Tool	What it does
mcmc_run	fit a model or spec and record the run
mcmc_diagnose	R-hat, ESS, MCSE, divergences, and a verdict
mcmc_summary	posterior mean, sd, and HDI per variable
mcmc_runs	the runs recorded in the project
mcmc_loo / mcmc_compare	out-of-sample fit, and a ranking
mcmc_sbc	simulation-based calibration
mcmc_doctor	whether a toolchain is ready

Each one runs the real command, so an assistant sees exactly what you would. A fit that runs but does not converge comes back as diagnostics rather than an error, because that is a result to act on, not a failure to retry.

Why the checks matter more here

A model written by an assistant is plausible by construction and correct only by luck: it can encode the wrong likelihood, an improper prior, or a parameterisation that samples badly. The useful loop is not “generate a model” but “generate a model, then prove it”.

1. mcmc run --prior and look at the prior predictive draws.
2. Fit, and read the diagnostics rather than the estimates.
3. mcmc sbc to test that the posterior is calibrated at all.

That is the part a chat window cannot do, and it is why the server exposes sbc and loo as first-class tools rather than only run.