A month ago, if an agent hit a sign-in page, it borrowed your browser. Your Chrome, on your machine, competing with the agent's own work for memory. You signed in, stepped away, and the session lived in the one place you were also using.
We moved that browser. Every octo, which is what we call an agent in Octomind, now has its own cloud browser running on our side, with its own profile, a location you pick, and a live view you can grab the keyboard in. I'm Don Karter, co-founder and CEO at Muvon, the team building Octomind. Here are the parts that matter if you're wiring agents to the web.
The session boundary is the part to think about
Octos used to ask for a sign-in up front. Now they open the site first and ask only when a login page, a 2FA prompt or a CAPTCHA actually appears. Sites that never needed a login stop interrupting you.
Where the password goes matters more. It travels from your keyboard to the site and never reaches the model. The octo does get the signed-in session, though, and a session is access. So give it the accounts you'd give a contractor: enough to do the job, nothing it doesn't need.
Each octo has its own browser profile, so sites you signed in to stay signed in the next time it opens one. Two octos don't share cookies. That's the boundary I'd design around: one agent, one profile, one blast radius. Separate octos for work and personal accounts keep those sessions apart.
Pages render as they do for you, scripts and all, so the agent sees what you'd see. The browser closes on its own when the octo is done with it.
Location is a setting, and it's sticky
Under the Setup tab, Browser location offers United States, United Kingdom, Germany, France, Singapore, or No fixed address. The five named ones give you the same country address every time. No fixed address gives a new address each time.
Fixed addresses help in two concrete cases: sites that challenge a login from a new place, and shops or search results that change by country. If your agent checks a price or a regional catalog, a moving address makes the results move with it.
A location change applies the next time the octo opens a browser, so set it before the task starts rather than during it.
Handover beats a workaround
While an octo is browsing, its Work tab says so, with a Watch or take over link. That opens the live browser in the panel. You can watch it click through, or take the mouse and keyboard yourself and hand it back when you're done.
For agent builders this kills a category of workaround. The weird 2FA flow, the page that only behaves for a human, the one field you don't want a model filling in: you take the controls for a few seconds instead of writing a special case for it.
What a browser minute costs
The cloud browser is part of the paid plans. It's $0.02 each time a browser starts, and $0.015 per minute it's open, with the last minute rounded up when it closes. A ten-minute errand on one site costs about 17 cents. On Free, octos don't get the browser, and signing in to sites needs a subscription.
Every start is $0.02 on top of the clock, so batching an agent's pages into one session costs less than opening a browser per step.
Spending is easier to read alongside that. The conversation header now shows Today $X · Total $Y, and clicking it opens Conversation spending: that conversation's model costs, day by day, newest first. Days are in UTC, and machine time and activation fees aren't included.
The rest of the update, where it touches agent work
Some tools need something only you know: which environment to deploy to, what size to order, whether to go ahead with a refund. An octo now asks in a form headed Your octo needs a few details, naming the tool that's asking. You Submit, or pick Don't share, or Not now. A request to open a link shows the site's address first and warns you to open it only if you trust it.
Follow-ups that are clearly part of the job you gave get done without a "shall I?", with at most one suggested next step at the end. Routines on a schedule now report what they did instead of ending on a question nobody is there to answer.
Paid plans run more model calls in parallel: Pro 3 to 5, Max 5 to 10, Team 15 to 20 shared. A call over the limit waits in line for up to 20 seconds, up from 15, before it's turned away. Machines run Octomind 0.67.0 from their next start, model calls travel over one persistent connection, and an octo reads several files at once instead of one by one.
What I'd do with this
Give each agent the narrowest account that does the job, pin its location to the country whose results you want, and watch it live the first few times it touches a site you care about. Then look at the per-minute number and decide whether that task should browse at all.
Full pricing and location details are in the cloud browser update.
Top comments (0)