Introduction

MCMC.js is a thin TypeScript command-line tool that automates the Bayesian workflow across probabilistic programming languages. You write a model in Turing.jl, JuliaBUGS, or Stan, and the mcmc CLI provisions the runtime, runs the sampler, and turns the result into convergence diagnostics, summary tables, and plots.

The CLI does not reimplement inference. It orchestrates: it owns argument parsing, spec validation, file I/O, diagnostics, and bootstrapping, and reaches into the backend’s runtime (Julia, or CmdStan for Stan) only to run the model itself.

One spec in, one samples file out

Every stage shares a single data contract.

• One declarative spec file in describes the model, its data, and the sampling configuration. TOML is the primary format; JSON is also accepted.
• One samples file out is the canonical output of inference and the canonical input to everything downstream. It is MCMCChains JSON, a cross-ecosystem container for posterior draws plus sampler statistics (divergences, energy, tree depth).

Because the output of one command is a valid input to the next, the commands compose. Choosing a cross-ecosystem samples format on purpose also insulates MCMC.js from Turing’s in-memory chain-type change: the Julia driver requests a stable chain type, and downstream commands never depend on Turing’s internal representation.

The workflow

model  ->  infer  ->  diagnose  ->  predict

• Model in a .jl or .stan file, a spec.toml, or a DoodleBUGS graph.json.
• Infer with mcmc run (zero-config) or mcmc fit (spec in, samples out).
• Diagnose with mcmc diagnose and mcmc summary, visualize with mcmc plot.
• Predict posterior-predictive draws with mcmc predict.

Who it is for

MCMC.js targets humans and AI agents at the same time, which drives two design choices.

• Structured output. Every command supports --json, so results are machine-parseable.
• Clear exit codes. 0 ok, 1 error, 2 ran but a domain check failed (for example, non-convergence). An agent can branch on the exit code without parsing text.

Next steps

Install the CLI and provision a runtime in Installation, then fit and diagnose your first model in the Quickstart.