From Graveyard to Glory

Production Python in the Browser

Mahmoud Hashemi

PyCon US 2026

A deep dive on
hand-building
a browser Python app,
byte by byte

JK

/plan i’d like to have a “python-in-browser” setup that lets me test regexes using python’s re module, for regex testing with python’s regex specifically. research and propose the best architecture for this in 2026. apply the color scheme #FDA500 for accent, #3FABCC for action (light) and #3F8BA3 for action (darker, for like hover). We want a dark mode, so invert and convert as necessary.

One prompt.
One plan.
$16 in tokens.

`yak.party/pyregex`

What Just Happened?

Talk over?

Have I been replaced by AI?

😂

Agents don’t give you the most important parts.

This Talk

What's coming up

The WHY

Neither syntax nor dogma.

The History

A decade of attempts. What changed.

Five Gaps

What the agent doesn’t know about production.

The Future

A glimpse at what’s coming.

Reflections

What this all adds up to.

Part One · of Four

The WHY

Not syntax. Not dogma.

Python is awesome!

(and JavaScript is terrible)

syntax isn’t a reason

Python is awesome!

(and JavaScript is terrible)

and neither is dogma or aesthetics

Why Browser Python?

§1 WHY

Actually good reasons:

Budget - free compute on every device
Sandboxing + Security - no server compromise
Reproducibility - same results, client-side
Python is Awesome!! - The ecosystem

Part Two · of Four

The Graveyard

A brief, honest history of Python in the browser

The Graveyard

Years	Project	Approach
2009-2015	PyJS / Pyjamas	Transpiler + Qt/GTK widgets
2010-2021	Silverlight / IronPython	Browser plugin
2013-2019	PyPy.js	Compile PyPy to JS
2013-2021	RapydScript	Python-like to JS compiler

Still Kicking

Years	Project	Approach
2009-present		JS reimplementation, still chasing Python 3
2014-present		Runtime transpiler, maintained (Python 3.14), no C extensions
2016-present	Transcrypt	Python-to-JS transpiler, active (v3.9.4)

Alive, but still niche.

What Was Missing?

Every previous attempt lacked at least one of:

Python syntax
Python stdlib
Python ecosystem (pip install)
DOM access
Bidirectional JavaScript interop

Then: WebAssembly + Pyodide

2011 - Empythoned (Replit ): Emscripten + CPython.

2017 - WebAssembly ships in all four browsers (overtaking Emscripten in Google Search results)

2019 - Pyodide: CPython compiled to WASM (Mozilla Research)

2020 - micropip: pip install in the browser (Pyodide 0.15)

2026 - v0.29. CPython Tier 3 WASM support (PEP 816)

Built on

Project	What	Link
JupyterLite	Full Jupyter in browser (official subproject, Feb 2026)	jupyterlite.readthedocs.io
Marimo	Reactive notebooks, WASM export	marimo.io
Gradio-Lite	ML demos in browser (Hugging Face)	huggingface.co
Stlite	Streamlit → WASM, PWA, Electron	github.com/whitphx/stlite

Common thread: Mostly directly about Python in the user’s browser.

Part Three · of Four

What the agent doesn't know

What we built (and what the agent didn't give us).

Production Python in the browser

A couple projects I’ve shipped.

glompad: yak.party/glompad

Side project: Data exploration with glom

FinFam: finfam.app

Work: Like GitHub for consumer finance.

NB “Python” isn’t the point, it’s just the best tool for this job.

What the Agent Built

The architecture

What’s Missing?

Five gaps to close

1 Testing Playwright, CI, and git hooks
2 UI Framework Svelte, compilers, and linters
3 Typed boundaries Comlink and worker RPC
4 Telemetry error reporting and closeable loops
5 Worker State Management state machine + interrupts

Gap 1: Playwright + CI

Browser Python apps need browser tests.

Playwright: E2E across 3 browsers
GitHub Actions: CI on every push
Pre-commit hooks: catch problems before they hit CI

Gap 2: UI Framework

Compilers, linters, and plugins the agent left out

Svelte: compiler catches errors at build time, not runtime
Plugins: TypeScript, ESLint, Prettier in the pipeline
Hook into pre-commit + CI (the agent didn’t wire these up)

Every layer of the build is a verification step.

Gap 3: Comlink

Typed boundaries are your friends.

What the agent builds:

// Manual postMessage RPC
const pending = new Map();
let nextId = 0;

function execute(code) {
  const id = nextId++;
  return new Promise((resolve, reject) => {
    pending.set(id, { resolve, reject });
    worker.postMessage({ id, type: 'exec', code });
  });
}

worker.onmessage = (e) => {
  const { id, result, error } = e.data;
  const p = pending.get(id);
  pending.delete(id);
  if (error) p.reject(error);
  else p.resolve(result);
};

What production uses:

// Comlink - 4KB
import { wrap } from 'comlink';

const api = wrap(
  new Worker('./worker.js')
);

// That's it. Call Python like a function.
const result = await api.execute(code);

Comlink turns postMessage into transparent function calls.

4KB. Zero boilerplate.

Gap 4: Telemetry

Closeable loops: The browser is an environment. Treat it like one.

Gap 5: Worker State Management

Interop and architecture

The agent gives you isReady and isLoading. Production needs a state machine:

SharedArrayBuffer enables fast interrupts like Ctrl-C (but requires COOP/COEP headers).

Part Four · of Four

The Future

A glimpse of what's coming.

What’s Coming

§4 THE FUTURE

What	Why It Matters
CPython Tier 3 WASM (PEP 816)	Official upstream support
emscripten-forge	Recreating conda in WASM
Cloudflare	Pyodide in V8 isolates. Web UX doesn’t stop at the browser.

Edge Computing

§4 THE FUTURE

server → edge → browser

Cloudflare Workers: Pyodide at the edge - ~1s cold starts via memory snapshots

Part Five · of Four

Reflections

What this all adds up to.

Ps and Qs

§5 REFLECTIONS

Verifiability: the thread that tied our five gaps together

Tests - Playwright E2E, not just unit tests
Types - Comlink typed boundaries
Compilers - Svelte, TypeScript, Rollup
Linters - ESLint, Prettier, CI checks

Close your loops so you can focus on creating.

Building in the age of plastic software

§5 REFLECTIONS

Python won. The language of good enough.

Good enough is glorious.

What hasn’t changed: Reach correctness first. Optimize later.

With every passing release, Python is becoming less and less the bottleneck.

The prototype is the product.

Go forth and build

§5 REFLECTIONS

Now is the time to build first-class apps around Python’s top-notch libraries.

Got a favorite Python library? Build a web application around it.

Beautiful, interactive web UX.

Rock-solid Python business logic.

Learning in the age of agents

Python may be the language agents are best at writing.

Python may have won, but what about us?

An agent-assisted human beats a pure agent. Every time.

Ecosystem and community are real advantages. Don’t let them disappear.

If you believe one good turn deserves another…

goodturn.ai

Links & Resources

Projects shown today:

glompad - Production browser-Python app
FinFam - Production browser-Python at work
PyRegex - State of agents in 2026
Pyodide - The runtime
GoodTurn - If one good turn deserves another

Learn more:

sedimental.org - Blog
github.com/mahmoud - Code
A Talk Near the Future of Python - dabeaz’s 2019 byte-by-byte version

From Graveyard to Glory

A deep dive on hand-building a browser Python app, byte by byte

JK

One prompt.One plan.$16 in tokens.

yak.party/pyregex

What Just Happened?

Talk over?

😂

This Talk

The WHY

Why Browser Python?

The Graveyard

The Graveyard

Still Kicking

What Was Missing?

Then: WebAssembly + Pyodide

Built on

What the agent doesn't know

Production Python in the browser

What the Agent Built

What’s Missing?

Gap 1: Playwright + CI

Gap 2: UI Framework

Gap 3: Comlink

Gap 4: Telemetry

Gap 5: Worker State Management

The Future

What’s Coming

Edge Computing

Reflections

Ps and Qs

Building in the age of plastic software

Go forth and build

Learning in the age of agents

Links & Resources

A deep dive on
hand-building
a browser Python app,
byte by byte

One prompt.
One plan.
$16 in tokens.

`yak.party/pyregex`