Browser

Operates like a human

A web product with no MCP server, operated the way a person would: a browser, on the sites you allow.

What a member can do

Some products have no MCP server: an old CRM, a supplier's portal, an internal admin page. A member can still use one, by driving a browser through it. You bind a browser to a prefix you choose, such as crm, with the sites it may open, and a task names the actions it may take there:

Action What it does
crm.navigate Opens a URL on one of the sites you allow.
crm.snapshot Reads the page as a list of its elements, each with a reference the other actions can take. With since: true, it returns only what changed since the last snapshot of the same page.
crm.extract Reads the text of an element: body for the whole page.
crm.click Clicks an element.
crm.type Fills a field with text.
crm.select Chooses an option in a dropdown, by the label the dropdown shows.
crm.press Presses a key, such as Enter, Escape or Tab.
crm.hover Moves the pointer over an element, so a menu that only opens on hover opens.
crm.scroll Brings an element into view, or scrolls the page.
crm.wait Waits until the page reaches a state: its address contains some text, some text is visible, or an element is visible, hidden or enabled. Up to three minutes.
crm.switch_tab Moves to another tab the site opened, such as a popup or a payment window.
crm.screenshot Captures the page as it is.
crm.download Takes a file the site offers and keeps it with the run.
crm.upload Attaches a file the run holds to a file field. Always waits for your approval.
crm.sign_in Stops at a sign-in page and waits for a person to sign in.

A task has to name each action it may take. The bare prefix, crm, gives it no browser at all, and one action grants no other: a task that lists crm.click cannot crm.press. What each action takes is in the Computer-Use tools reference.

An action finds its element by a reference from the latest snapshot, such as e12, or by a CSS selector. A reference works inside an iframe too, and only until the page changes: after that the member takes a new snapshot.

Each run gets its own browser, a headless Chromium, opened at the start URL on its first action and closed, with every tab in it, when the run ends. It keeps nothing from an earlier run, except a sign-in a person completed for that member on that site (see Signing in). A browser is only ever driven by a run of a task: a member's chat never offers one.

Only the sites you allow

Every page the browser opens has to be on one of the binding's sites: the host itself, or any subdomain of it. A navigation anywhere else is refused before the browser moves, and the run is told which site was blocked (computer_use_site_blocked). Where the browser actually lands is checked again after every action, so a link or a redirect cannot take it off the list either: it is taken back to the last allowed page, or closed if it cannot be, and the action is refused. A tab the site opens on another host is left open, and the member cannot switch to it.

The sites are a rule about where the browser goes, not about what a page sends. A page on an allowed site can still send requests to other hosts, such as an embedded payment form posting to its processor, or analytics, and those go out as any browser sends them.

Files in and out

crm.download takes a file by its address or by the control that starts the download. The address, and every redirect on the way, has to be on the binding's sites. The file is kept with the run, and the member is told its name, size and type, never its contents.

crm.upload attaches only a file the run already holds: one it downloaded, or one the run was started with. A task can declare a file input, and Start a run then shows a file picker for it. A path on a computer or a URL is refused. Every upload waits for you to approve the file and the site first. Rejecting it cancels the run, as rejecting any gate does, and the file is never attached.

What the model reads

Anything typed into a field that looks like it holds a secret (a password, a token, a key, a PIN), and every cookie the browser holds, is taken out of everything the member reads back: results, addresses, and error messages alike. A screenshot is stored with the run and never put in front of the model: the member reads a page with snapshot or extract.

Which models can drive it

A run whose task names a browser action starts only on a model trusted to operate a browser. Today those are:

  • Anthropic: Claude Opus, Sonnet, Haiku and Fable from version 4 on, and Claude 3.5 Sonnet.
  • OpenAI: GPT from 5.4 on, including the 5.6 and 6 families and GPT-6.1 Sol. Not GPT-5.4 nano, and not a Pro variant.
  • xAI: Grok from 4.7 on.

Each counts on its own provider or through OpenRouter. On any other model the run stops before the model is called, with computer_use_model_unsupported, and shows on Blockers: change the member's model, then retry it. See Spend and models.

Who connects it

A browser binding holds no credential. It is three things: a prefix, the sites, and a start URL on one of those sites. What the sites see is a signed-out visitor, until a person signs in for the member.

A binding can sit on the enterprise, a team, a member, a role or a task, and the one further down replaces the one above: task, then role, then member, then team, then enterprise. A binding lower down may leave out sites that a wider one it inherits from allows. Adding a site that the wider one does not allow is refused (widens_<layer>_grant), unless you are the person who bound the wider one, or the CEO.

Signing in

A member never types a username or a password itself. When its browser reaches a sign-in page, it calls crm.sign_in, which its task has to list like any other action, and the run stops.

The run's page then shows Waiting on a sign-in, with the site and the page the browser is on. Choose Take the browser, sign in by naming a field or a button and typing or clicking, then choose Hand back. What you type is sent to the page only: the run's record and its model never see it. Only someone who may decide the run's gates can hand the browser back. The run carries on from the page you left it on.

What the sign-in left behind (the site's cookies and stored session, never a password) is kept for that member on that site, so the member's later runs start signed in. It is never shared with another member, and a copied task does not carry it. When the site has ended that session, the next run meets the sign-in page and asks again.

Each browser binding on Tools lists the sign-ins it keeps, by member and site, with Revoke sign-in beside each: the next run for that member on that site meets the sign-in page again.

What waits for you

Nothing a browser does waits for you unless the task declares a gate on it, with two exceptions: crm.upload always waits for you to approve the file, and crm.sign_in waits for a person to sign in.

Gate one action, such as crm.click, and every click waits; gate the prefix, crm, and every action does, reading included. A gated action waits with exactly what it was about to do (the selector, the text, the address), and nothing happens on the site until you approve it on Gates. Approving runs the action together with whatever it sends, and no second approval follows. A gate on one action holds no other: a gate on crm.click does not hold a crm.press of Enter that submits the same form. See Gates and tool scopes.

An approval waits until someone decides it or its run is cancelled. A sign-in is different: one that nobody completes before your enterprise's gate timeout cancels the run.

What a run does without asking

On a site its binding allows, a run submits, pays, sends and deletes without asking anyone, unless the task gates the action. What an action sets off goes to the site as a browser sends it: a form that posts, a save in the background. So does what the page sends on its own between actions. None of it is held or approved separately.

If you want to see a submission before it happens, gate every action that can submit it: click, type, press and select, or the prefix. Or keep the binding's sites to pages where nothing can be saved, sent or deleted.

Setup

  1. In the portal, on Tools (or a member's or team's page → Tools), choose Add tool.
  2. Kind: Computer-Use. Tool type: the prefix your tasks will use, such as crm.
  3. Allowed sites: one host per line, with no https:// and no path: crm.example.com. A host covers its subdomains. List only what the work needs.
  4. Start URL: where the browser opens, on one of those sites: https://crm.example.com/.
  5. Choose Bind browser.
  6. On the task, name each action it needs (crm.navigate, crm.snapshot, crm.click, …), with crm.sign_in when the site needs a sign-in, and declare a gate on each one that should wait for you.
  7. Check it. Run the task. The run page shows every action, and the address the browser was on after it.

Troubleshooting

What you see Why What to do
computer_use_not_on_task The task lists the bare prefix, crm, not its actions Name each action: crm.navigate, crm.click, …
computer_use_site_blocked: <host> The page is on a host the binding does not allow Add the host to the binding's sites, if the work needs it
computer_use_model_unsupported The run's model cannot operate a browser Change the member's model to one that can, then retry the run
computer_use_browser_unavailable No browser could be started for the run Nothing about your task is wrong. Retry the run from Blockers
Calls fail as though nothing were bound The start URL is not on one of the binding's sites, so it gives no browser Put the start URL's host in the sites
Saving a binding is refused when it adds a site (widens_<layer>_grant over the API) A wider binding it inherits from does not allow that site Ask whoever bound the wider one to add it there, or leave the site out
computer_use_stale_ref The reference is from an earlier snapshot, or the page has changed since Nothing was done. The member takes a new snapshot and uses a reference from it
computer_use_snapshot_too_large The page is too large to read in one snapshot Have the task snapshot one part of the page, fewer levels deep, or only what changed
computer_use_wait_timeout The page did not reach the state a wait asked for in time The browser stays open and the run carries on. Check that the task waits for something the page really shows
computer_use_unknown_option A select named a label the dropdown does not offer The error lists the labels it does offer. The label has to match exactly
computer_use_tab_not_found A switch_tab matched no open tab The error lists the open tabs, by title and host
computer_use_upload_refused An upload named something that is not a file the run holds Have the task upload a file it downloaded, or one the run was started with
computer_use_sign_in_not_completed Nobody signed in before the enterprise's gate timeout, so the run was cancelled Start the run again, and sign in from the run's page when it asks
A click or a field is not found The selector does not match the page as it is Have the task snapshot the page to see what is there, and act by reference