Your agent compares pricing pages, pulls numbers from your dashboards, and fills in forms, in a browser you can watch.
You'll need the bb desktop app for sites you're logged in to.
Ask me about each of these, one at a time, with a suggested default: - Task (what you want found, compared, or filled in) - Sites (where to look) - Logins (sites that need your account, if any) - Bring back (e.g. a table with a source link per row) Skip any I've answered. Then sum up the plan in three bullets and wait for my go. My answers override the steps below. If a wait times out, wait again, up to three times. Use a real browser to do the web task above, and show me what you found. Guide: https://getbb.app/guides/agent-browser-for-work Do every step below, with a Check after each step. Stop if a check fails and tell me what failed. 1. Set up browser access. Run bb plugin list. Browser Automation isn't installed by default: install it with bb plugin install browser-automation --yes if it's missing, or enable it with bb plugin enable browser-automation if it's off. Agent Annotations is built in but off: enable it with bb plugin enable agent-annotations. Then read bb guide browser and bb browser-automation --help. Check: both plugins are running. 2. Open a browser. Find the computer running the bb desktop app with bb machine list, then run bb browser instances --host <desktop-host-id> --json. If a desktop instance is connected, open a Browser tab there so I can watch and my sign-ins work: bb browser-automation open --backend desktop --machine <desktop-host-id> --desktop <instance-id> --json If no desktop instance is connected and Logins lists any site, stop and tell me signed-in sites need the bb desktop app open. Otherwise open headless Chrome on this thread's machine: bb browser-automation open --backend local --headless --machine <host-id> --json Keep the returned session ID. Emit the returned previewDirective once if present. Check: the first page loads. For a signed-in site, confirm the expected account is signed in. If it isn't, stop and ask me to sign in once in a Browser tab, or to import my sign-ins in Settings → Browser. Never read or print cookie values. 3. Do the task. Take a fresh snapshot before each action and use its element references. Read pages instead of guessing, and keep a list of the sources you used. If a page doesn't state a number, say so instead of filling it in from memory. Don't buy anything, send messages, submit forms that commit me to something, or change account settings unless I asked for exactly that; stop and ask first. Check: you have the information or result I asked for, with the page each fact came from. 4. Report and leave the tab for feedback. If you used a desktop Browser tab, stop controlling it with bb browser-automation stop <session-id> so it stays open on the most relevant page, and tell me I can choose Annotate elements to point at anything to change or dig into. If you used headless Chrome, close it with bb browser-automation close <session-id> and give me the page URLs instead. Check: reply with the result in the format from my Bring back answer (by default, a table with a source per row), a source link for each fact, screenshots of anything I should look at, and what you couldn't do. If it's a task I'll want again, offer to save these steps as a bb skill in .bb/skills/<short-name>/SKILL.md with my answers filled in, or as a bb automation if it should run on a schedule.
The Browser Automation plugin gives your agent a browser. The Agent Annotations plugin lets you point at what you mean.
For dashboards and tools behind a login, start bb's Browser signed in.
Say what you need, which sites to use, and what to bring back.
Choose New thread, set the project to Don't work in a project, then paste your request:
Compare the cheapest paid plans of Linear, Trello, and Asana from their pricing pages: the plan name and its price per user per month, billed yearly. Work in a Browser tab here so I can watch. Put the result in a table and link each price to the page it came from.
Jobs that work well:
While your agent uses the tab, a bar shows it's in control. Choose Take over to click or type yourself, like for a two-factor code, or Stop to end its control.
Select the thing on the page instead of describing it.
Your agent closed its browser, or it sat unused for five minutes. The thread keeps the last view. Ask your agent to open it again.
No. When you start the thread, set the project to Don't work in a project, and ask for the job.
Not unless you ask. The prompt tells it to stop and ask first, and for a hard stop, choose Accept Edits in the permission menu below the message box. Your agent then needs your OK before it acts in the browser.
Ask your agent to say so instead of guessing, and to link the page it checked so you can look yourself.
Claude Code, Codex, and Pi are built in. Cursor, opencode, and other agents that support the Agent Client Protocol, like Grok Build and Hermes Agent, work once they're installed.
Not for a fresh browser: your agent can open Chrome on any connected Mac or Linux machine, and you watch it in the thread. Browser tabs, your logins, and annotations need the desktop app.
Yes. A Browser tab shows a bar while your agent controls it. Choose Take over to click and type yourself, like to enter a sign-in code, or Stop to end its control.
No. bb imports your logins, not your saved passwords, so your agent opens sites already signed in without seeing a password. Your logins stay in bb's Browser on this computer, and the prompt from Copy for agent tells your agent never to read or print them.
Yes. Ask for the browser on the machine you want. There, localhost means that computer, so use an address it can reach.
No. Browser Automation talks to Chrome and bb's Browser directly, so it doesn't need Accessibility or Automation permission.
Free and open source. Use the agents you already have.