stan web IDE#

Run Stan sampling in your browser, inside a VS Code-style IDE built on minwebide. Projects live in your browser's IndexedDB. Models compile on a remote stan-wasm-wasi server to pure WASI modules; sampling itself runs locally, one web worker per chain, each invoking the module CLI-style through a small WASI shim.

Live site: https://concept-collection.github.io/stan-web-ide/

How it works#

A project holds .stan programs, .json data files, and .sample files. A .sample file is a YAML description of one sampling run:

stan: linear.stan      # the Stan program
data: data.json        # the data
num_chains: 4          # optional; defaults 4 / 1000 / 1000 / 2.0 / random
num_warmup: 1000
num_samples: 1000
init_radius: 2.0
seed: 42               # omit for a random seed

Paths are relative to the .sample file (leading / = project root). Opening a .sample file shows a form view — file pickers, sampling parameters, a Run button, and per-chain progress bars. The form edits the underlying YAML (tab menu → Reopen as Text Editor for the raw file); the tab bar's ▶/⏹ runs and stops the same way. Runs use current editor contents, saved or not.

A run compiles the program (server-side, cached by source hash), streams Stan's console output to the Output panel, and writes into the output directory — always <name>.out next to the .sample file (so fit.sample → fit.out, replaced on each run):

• chain_1.csv … one CSV per chain, header = parameter names, one row per draw
• summary.csv — mean, MCSE, sd, 5%/50%/95%, ESS, ESS/s, split-Rhat per parameter (via mcmc-stats)
• console.txt — the sampler's console output
• run.json — the exact configuration used (including the resolved seed); written last, so it doubles as the run's completion marker

When a run finishes, a results dashboard opens on run.json: tabs for the summary table (Rhat highlighted when > 1.01), histograms, trace plots, scatter plots, a draws table, and the console — plots via lazily-loaded plotly. It reads the sibling CSVs, so it reopens in any later session: click View results on the .sample form, or run.json in the output folder.

The .stan editor has syntax highlighting plus diagnostics, hover docs, completion, and auto-format from stan-language-server (stanc3 compiled to JS, running in a worker).

The compilation server#

Compiling Stan to WebAssembly needs a server; everything else is local. The status bar shows the configured server (click it to change; persisted in the browser). The default is https://stan-wasm-wasi.fly.dev, a hosted stan-wasm-wasi instance — it allows any origin (CORS), caches compiled models by source hash, and auto-stops when idle, so the first compile after an idle period pays a ~30 s cold start. To run one locally instead:

docker build -t stan-wasm-wasi https://github.com/magland/stan-wasm-wasi.git
docker run --rm -p 8083:8080 stan-wasm-wasi

then set the server URL to http://localhost:8083.

The compiled models are pure-WASI command modules (main.wasm): each run of the module executes one MCMC chain (CmdStan seed/chain-id convention), with draws on stdout as CSV and Stan's console on stderr. Chains run in parallel workers, so no threads, no SharedArrayBuffer, and no cross-origin isolation are needed.

Development#

minwebide is consumed as a sibling checkout (file:../minwebide):

git clone https://github.com/magland/minwebide ../minwebide
(cd ../minwebide && npm install)   # fetches the pinned VS Code source
npm install
npm run dev                        # http://127.0.0.1:3000

• npm run build — static bundle in dist/
• npm run typecheck — typechecks app code (vendor diagnostics suppressed)
• npm run smoke — headless end-to-end test against the built bundle; when the default compile server is reachable it also compiles and samples for real
• node scripts/dev-check.mjs — quick checks against a running dev server

CI checks out magland/minwebide next to this repo, installs both, builds, and publishes dist/ to GitHub Pages.