MCMC.js
Menu
Docs Install Quickstart GitHub

Bayesian inference, from one command line.

Install MCMC.js Read the docs
mcmc — zsh

mcmc drives Bayesian modelling end to end: provision a runtime, fit a model, and run post-inference diagnostics. Stages compose through one declarative spec file in and one samples file out, so each step stays small and predictable.

Eight schools, end to end.

The textbook hierarchical model, and a textbook trap. Run it, and the diagnostics flag the funnel the centered parameterization hides: high R-hat, low ESS, hundreds of divergences. Drag to zoom any plot. 

using Turing

@model function eight_schools(y, sigma)
    J = length(y)
    mu ~ Normal(0, 5)
    tau ~ truncated(Cauchy(0, 5); lower = 0)
    theta ~ filldist(Normal(mu, tau), J)
    for j in 1:J
        y[j] ~ Normal(theta[j], sigma[j])
    end
end

build_model(data) = eight_schools(data.y, data.sigma)
y,sigma
28,15
8,10
-3,16
7,11
-1,9
1,11
18,10
12,18
Run it 
$ mcmc run eight_schools.jl --data data.csv
Fitting eight_schools.jl with Turing.jl (NUTS, 4 chains x 1000 draws + 1000 warmup) on Julia 1.12.6...
ok: run 20260626-062103-35158c (19.8 s, Julia 1.12.6) saved to .mcmc

variable  mean   std    r_hat  ess_bulk  ess_tail  mcse   hdi
--------  -----  -----  -----  --------  --------  -----  ----------------
mu        4.748  3.219  1.042  114       1250      0.319  [-1.604, 10.218]
tau       3.436  3.375  1.359  17        15        0.498  [0.310, 9.348]
theta[1]  6.447  5.424  1.186  578       983       0.198  [-3.527, 16.557]
theta[2]  5.339  4.501  1.101  350       1279      0.202  [-3.439, 13.942]
theta[3]  4.312  5.244  1.047  153       1185      0.383  [-6.738, 12.959]
theta[4]  5.192  4.664  1.094  304       1534      0.239  [-3.846, 14.176]
theta[5]  3.962  4.562  1.043  110       1208      0.449  [-5.701, 11.438]
theta[6]  4.378  4.658  1.082  190       1622      0.325  [-4.396, 13.254]
theta[7]  6.530  4.773  1.188  579       675       0.183  [-2.170, 15.820]
theta[8]  5.309  5.205  1.089  338       1464      0.180  [-4.724, 15.047]

divergences: 503
not converged (R-hat <= 1.01, ESS >= 400, divergences <= 0)

A thin, composable workflow.

mcmc orchestrates probabilistic programming languages without becoming a framework. One spec in, one samples file out.

Fit anywhere

Run Turing, JuliaBUGS, or Stan models through a managed Julia or CmdStan runtime, with the CLI as a thin TypeScript orchestrator.

Diagnostics built in

Compute R-hat, ESS, MCSE, and HDI over your samples to judge convergence with confidence.

Reproducible runs

A declarative spec file in and a samples file out keep every run inspectable and repeatable.

Start in one line.

Install the CLI from npm, then provision the runtime with mcmc setup.

npm global install Installs the mcmc binary. Run mcmc setup once to provision Julia via juliaup, or mcmc setup --engine stan for CmdStan. 
npm i -g mcmcjs
Quickstart

Documentation.

Focused pages for installation, running inference, diagnostics, the CLI reference, and internals.

Installation

Install the CLI and provision a runtime with a single setup command.

Run inference

Fit a model from a spec file and produce a samples file you can diagnose and plot.

Diagnose convergence

Read R-hat, ESS, MCSE, and HDI to judge whether your chains have converged.

FAQ

Quick answers for search engines and humans.

What is MCMC.js?

Command-line tools for Bayesian modelling, MCMC inference, and post-inference diagnostics across PPLs.

Does it run the inference?

The CLI is a thin orchestrator. Inference runs in a Julia or CmdStan subprocess; diagnostics are pure TypeScript.

How do I check convergence?

Run mcmc diagnose to compute R-hat, ESS, MCSE, and HDI over your samples.