Automate any browser, behind one API

webtasks runs Chrome on your server and turns each automation flow into a typed HTTP endpoint. Write a short .webtask recipe — go here, wait, click, extract — and call it with plain JSON from any language. No headless browser in your app, no Selenium grid to babysit.

1 static binary REST + live SSE 38 ready examples GPL-3.0

Install in 10 seconds See it in action Star on GitHub

---

The problem it solves

Browser automation usually means bolting a headless browser, a matching driver, and brittle scripts into every service that needs it. That's heavy to run, hard to secure, and a nightmare to keep in sync across a fleet.

webtasks flips it around: one automation server, called over HTTP. Selectors, login flows, and Chrome internals live in one place. Everything else just sends JSON and gets JSON back.

The old way	With webtasks
A headless browser + driver in every service	One server, called over HTTP from anywhere
Selenium grid / chromedriver version matching	A single static binary talks to Chrome via CDP
Selectors & logins copy-pasted across repos	Recipes live in one bundle, reused everywhere
Rebuild & redeploy to change a flow	Edit a .webtask file — hot-reloads on next call
Brittle, undocumented scripts	A typed HTTP API with input/output schemas

---

A task is a recipe

Drop this in tasks/crawl/hackernews-top.webtask and it instantly becomes POST /tasks/crawl/hackernews-top:

task "crawl/hackernews-top"
    pool default
    timeout 20000
    transport rest

    goto "https://news.ycombinator.com"
    wait until "tr.athing" timeout 10000

    extract stories from "tr.athing" repeat
        title text ".titleline > a"
        url   attr href on ".titleline > a"
    end
end

Call it from anything that speaks HTTP:

curlPythonJavaScript

curl -s -X POST localhost:8765/tasks/crawl/hackernews-top -d '{}'

import requests
r = requests.post("http://localhost:8765/tasks/crawl/hackernews-top", json={})
print(r.json()["data"]["stories"])

const r = await fetch("http://localhost:8765/tasks/crawl/hackernews-top", {
  method: "POST", body: "{}",
});
console.log((await r.json()).data.stories);

{ "ok": true, "data": { "stories": [ { "title": "Show HN: …", "url": "https://…" } ] } }

---

What you can build

• Scrape & extract

  ---

  Pull structured JSON from any page or list with CSS-selector field specs. Real sites, real data.

• Drive UIs

  ---

  Fill forms, click, type, scroll infinite feeds, and wait for dynamic SPA state — with native, trusted input events.

• Capture artifacts

  ---

  Screenshots, full-page PDFs, MHTML archives, and animated GIF / MP4 recordings of a whole flow.

• Stream progress

  ---

  Long jobs emit live status and progress events over Server-Sent Events — perfect for progress bars.

• Inspect the network

  ---

  HAR-style request capture, cookie read/write, console logs, and network-idle waits for flaky SPAs.

• Stay logged in

  ---

  Persistent Chrome profiles + declared secrets keep authenticated sessions alive across restarts.

---

Built for developers

• Language-agnostic

  ---

  It's just HTTP + JSON. Call tasks from Python, JS, Go, shell — anything. GET /tasks returns the input/output schema for every endpoint.

• Readable recipes

  ---

  The .webtask language reads like a checklist, not a config file. No indentation traps, no boilerplate.

• Instant feedback

  ---

  Hot-reload re-reads recipes on every request. Edit, re-call, done — no restart, no rebuild.

• 38 examples to copy

  ---

  A demo bundle spanning scraping, forms, rendering, recording, and a real-world logged-in scrape.

Write your first task

---

Ready for production

• One binary, zero runtime deps

  ---

  A single static binary — no JVM, no chromedriver, no Selenium server. Ship it plus a zipped bundle and run anywhere Chrome is installed.

• Isolated window pools

  ---

  Concurrency is bounded per pool; a window is never shared by two runs at once. Crashed tabs are detected and replaced automatically.

• Secrets, never inline

  ---

  Credentials are declared in the bundle and resolved at startup from env, flags, or a prompt — surfaced to recipes as {{TOKEN}}, never hard-coded.

• Hardened by default

  ---

  Static file mounts are path-traversal–safe; per-call deadlines stop runaway runs; the server binds to localhost unless you opt out.

• Observable

  ---

  GET /health reports live pool occupancy and task counts; SSE streams every step as it happens.

• Portable bundles

  ---

  Config is a directory or .zip, loaded at runtime. The same binary serves any deployment — point it at a different bundle to change behaviour.

Deployment guide

---

How a request flows

flowchart LR
    Client["HTTP client<br/>curl · Python · JS"]
    Server["webtasks server"]
    Pool["window pool<br/>(bounded concurrency)"]
    Chrome["Chrome window<br/>chromedp / CDP"]
    Web["target site"]

    Client -->|"POST /tasks/name + JSON"| Server
    Server -->|"lease"| Pool
    Pool --> Chrome
    Chrome --> Web
    Server -->|"JSON or live SSE"| Client

Each .webtask file is one endpoint. A run leases a Chrome window for its whole duration and releases it at the end — so concurrency, sessions, and crash recovery are all handled for you.

How it works in depth

---

Get started

• Install

  ---

  curl … | sh, start the server, run your first task in under a minute.

• Writing tasks

  ---

  The complete .webtask language reference, with a build-from-scratch walkthrough.

• Examples

  ---

  38 runnable recipes across 11 categories — copy, tweak, ship.

• Deployment

  ---

  Pools, secrets, static mounts, and packaging a bundle for production.

---

webtasks is free software under the GNU General Public License v3.0. See License.