Skip to main content
When a task needs the web, Bed of Roses opens a real Chrome window and uses it the way you would — except it reads each page’s structure directly instead of squinting at pixels, so it’s fast and precise. You see both halves at once: the Chrome window doing the work on one side, and inside the thought bubble a browser card narrating each step — the page it’s on, the element it’s about to click, the fresh state after every action.

​
How it browses

  • A real, visible browser. BOR launches Chrome (or the bundled Chromium) headed by default, with its own persistent profile — separate from your personal browser, so its logins and cookies don’t touch yours.
  • It reads pages structurally. After every action, BOR gets an accessibility snapshot of the page — every link, button, and field, precisely identified. It acts on elements by reference, not by guessing coordinates.
  • It does real work. Clicking, typing, key presses, dropdowns, multi-field form fills, tab management, going back, waiting for content, and taking screenshots when it needs to actually look at something.
  • It can stay signed in. Because the profile persists, you can log BOR’s browser into your sites once (use “Open sign-in browser” from the connectors gallery) and it stays logged in for future tasks.
  • Verified before claimed. BOR re-reads the page after acting, so “I’ve submitted the form” means the form was actually submitted — not just that a click was sent.

​
What to say

  • “Find the cheapest direct flight to Lisbon in September.”
  • “Log into my electricity account and tell me the last bill.”
  • “Fill out this signup form with my details.”
  • “Compare the pricing pages of these three tools.”
One rule BOR follows: if a connector covers the service — Gmail, Calendar, Notion, GitHub — it uses the connector, never the browser. The browser is for the open web.

​
Quick lookups: search and fetch

Not everything needs a browser session. BOR also has:
  • Web search — ranked results with sources, rendered as a card in the bubble. This is how it gives sourced answers and researches before building.
  • Web fetch — read a specific URL’s content directly.
“What’s AAPL doing today?” is a search and a concise answer, not a browser session.

​
Fallbacks

  • browser-use CLI (optional). If you’ve installed the browser-use CLI, BOR detects it and can use it as a fallback backend — it attaches to your actual running Chrome (your real profile, cookies, and tabs) over CDP. This is entirely optional; nothing in BOR requires it, and without it the ambient browser handles everything.
  • Computer use. For the rare page that defeats structural reading, BOR can fall back to driving a visible browser with screen capture and clicks.

​
The Browser app

Separate from all of the above, BOR can open a plain Chrome window for you to use — the user-facing Browser app, with its own profile, distinct from the automation profile. Ask “open a browser” when you just want to look at something yourself.

​
Configuration

Environment overrides, if you need them: BOR_BROWSER_HEADLESS=1 (run the ambient browser headless), BOR_BROWSER_EXECUTABLE_PATH (a specific Chrome/Chromium binary), BOR_BROWSER_CHANNEL=chrome (browser channel), BOR_BROWSER_VIEWPORT (viewport size, default 1440x1000). npm run browser:install installs the bundled Chromium if you don’t have Chrome.