Plain-English tasks
Describe the outcome you want — fill a form, scrape a table, walk a multi-step flow — without writing selectors by hand.
Open source · Experimental · v0.6.0
Browsergent is Claude Code for the browser — an agent in your side panel that sees the page, writes JavaScript, and acts until the task is done.
BYOK · Anthropic-compatible · Chrome MV3 · MIT / Apache-2.0
An AI browser agent uses a large language model to operate a web browser the way a person would: read the page, choose the next action, click, type, navigate, and recover when something fails. Browsergent is a Chrome extension AI agent built for that loop — not a hosted RPA cloud and not a headless script runner.
You keep using your normal Chrome profile. The agent runs in the side
panel, observes the active tab, and drives it with a single tool:
run_js. That design keeps model reasoning separate from
browser side effects, so every action stays typed, inspectable, and
local.
Unlike remote browser farms, Browsergent operates on the page you
already have open — including authenticated sessions — through a
sandboxed page.* protocol.
Describe the outcome you want — fill a form, scrape a table, walk a multi-step flow — without writing selectors by hand.
Structured snapshots expose refs, roles, and labels so the model can ground actions in the live DOM.
The model never touches Chrome APIs directly. JavaScript runs in a sandbox that only emits typed page.* commands.
Use Anthropic, DeepSeek, GLM, or any Anthropic Messages–compatible endpoint. Your key stays in the browser.
Activate reusable skills and attach session files so repeated workflows stay consistent across runs.
No remote browser farm. Browsergent drives the tab you already have open — cookies, login, and all.
Walk sign-up, checkout, or admin wizards that break brittle scripts when the UI shifts.
Read pages, follow links, and pull structured facts from sites that don’t offer clean APIs.
Have the agent click through critical paths and report what broke — on your real browser profile.
Automate ops work inside authenticated SaaS UIs where traditional RPA is too rigid.
| Approach | You write | Runs on | Best for |
|---|---|---|---|
| Browsergent | Plain-English goals | Your Chrome tab | Changing UIs, exploratory flows, logged-in work |
| Playwright / Selenium | Selectors & scripts | Headless / remote browsers | Stable CI, deterministic test suites |
| Hosted browser agents | Prompts + cloud config | Vendor browser farm | Teams that want managed infra |
Browsergent is the right fit when you want open-source browser automation with an LLM in the loop, local control, and a BYOK model — not when you need a sealed enterprise RPA platform.
Browsergent is an open-source AI browser agent for Chrome. It lives in a side panel, reads the active page, and acts on it by generating JavaScript that runs through a sandboxed page.* command protocol.
Playwright and Selenium are scripted automation frameworks. Browsergent is an LLM agent: you describe the goal, it decides the steps, recovers from errors, and works on your real logged-in Chrome tab instead of a headless farm.
The extension runs locally. Model calls go only to the API base URL you configure (for example Anthropic). There is no Browsergent cloud that brokers your session.
Yes — MIT OR Apache-2.0. You pay only for the LLM provider key you bring (BYOK).
git clone https://github.com/Irvingouj/Browsergent.git
cd Browsergent && npm install && npm run build
Then load dist/ in chrome://extensions with
Developer mode on — or grab the
prebuilt zip.
Free, open source, and ready for Chrome in a few minutes.