[{"data":1,"prerenderedAt":48},["ShallowReactive",2],{"$f2ud55fvbsqk7l":3},{"href":4,"title":5,"description":6,"kind":7,"mark":8,"planned":9,"contributors":10,"provenance":8,"express":9,"bridge":9,"html":11,"headings":12},"\u002Fdocs\u002Ftools\u002Fbrowser","Browser","A web product with no MCP server, operated the way a person would: a browser, on the sites you allow.","operates",null,false,[],"\u003Ch2 id=\"what-a-member-can-do\">What a member can do\u003C\u002Fh2>\n\u003Cp>Some products have no MCP server: an old CRM, a supplier's portal, an internal admin page. A member can still use\none, by driving a browser through it. You bind a browser to a prefix you choose, such as \u003Ccode>crm\u003C\u002Fcode>, with the sites it\nmay open, and a task names the actions it may take there:\u003C\u002Fp>\n\u003Cdiv class=\"prose__table\">\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Action\u003C\u002Fth>\n\u003Cth>What it does\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.navigate\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Opens a URL on one of the sites you allow.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.snapshot\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Reads the page as a list of its elements, each with a reference the other actions can take. With \u003Ccode>since: true\u003C\u002Fcode>, it returns only what changed since the last snapshot of the same page.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.extract\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Reads the text of an element: \u003Ccode>body\u003C\u002Fcode> for the whole page.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.click\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Clicks an element.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.type\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Fills a field with text.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.select\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Chooses an option in a dropdown, by the label the dropdown shows.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.press\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Presses a key, such as Enter, Escape or Tab.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.hover\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Moves the pointer over an element, so a menu that only opens on hover opens.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.scroll\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Brings an element into view, or scrolls the page.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.wait\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>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.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.switch_tab\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Moves to another tab the site opened, such as a popup or a payment window.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.screenshot\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Captures the page as it is.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.download\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Takes a file the site offers and keeps it with the run.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.upload\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Attaches a file the run holds to a file field. Always waits for your approval.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>crm.sign_in\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Stops at a sign-in page and waits for a person to sign in.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\u003Cp>A task has to name each action it may take. The bare prefix, \u003Ccode>crm\u003C\u002Fcode>, gives it no browser at all, and one action\ngrants no other: a task that lists \u003Ccode>crm.click\u003C\u002Fcode> cannot \u003Ccode>crm.press\u003C\u002Fcode>. What each action takes is in the\n\u003Ca href=\"\u002Fdocs\u002Ftools\u002Fbuilt-in\u002Fcomputer-use\">Computer-Use tools reference\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>An action finds its element by a reference from the latest snapshot, such as \u003Ccode>e12\u003C\u002Fcode>, or by a CSS selector. A\nreference works inside an iframe too, and only until the page changes: after that the member takes a new snapshot.\u003C\u002Fp>\n\u003Cp>Each run gets its own browser, a headless Chromium, opened at the start URL on its first action and closed, with\nevery tab in it, when the run ends. It keeps nothing from an earlier run, except a sign-in a person completed for\nthat member on that site (see \u003Ca href=\"#signing-in\">Signing in\u003C\u002Fa>). A browser is only ever driven by a run of a task: a\nmember's \u003Ca href=\"\u002Fdocs\u002Fwork\u002Fchat\">chat\u003C\u002Fa> never offers one.\u003C\u002Fp>\n\u003Ch3 id=\"only-the-sites-you-allow\">Only the sites you allow\u003C\u002Fh3>\n\u003Cp>Every page the browser opens has to be on one of the binding's sites: the host itself, or any subdomain of it. A\nnavigation anywhere else is refused before the browser moves, and the run is told which site was blocked\n(\u003Ccode>computer_use_site_blocked\u003C\u002Fcode>). Where the browser actually lands is checked again after every action, so a link or\na redirect cannot take it off the list either: it is taken back to the last allowed page, or closed if it cannot\nbe, and the action is refused. A tab the site opens on another host is left open, and the member cannot switch to\nit.\u003C\u002Fp>\n\u003Cp>The sites are a rule about where the browser goes, not about what a page sends. A page on an allowed site can still\nsend requests to other hosts, such as an embedded payment form posting to its processor, or analytics, and those go\nout as any browser sends them.\u003C\u002Fp>\n\u003Ch3 id=\"files-in-and-out\">Files in and out\u003C\u002Fh3>\n\u003Cp>\u003Ccode>crm.download\u003C\u002Fcode> takes a file by its address or by the control that starts the download. The address, and every\nredirect on the way, has to be on the binding's sites. The file is kept with the run, and the member is told its\nname, size and type, never its contents.\u003C\u002Fp>\n\u003Cp>\u003Ccode>crm.upload\u003C\u002Fcode> attaches only a file the run already holds: one it downloaded, or one the run was started with. A task\ncan declare a \u003Ccode>file\u003C\u002Fcode> input, and \u003Cstrong>Start a run\u003C\u002Fstrong> then shows a file picker for it. A path on a computer or a URL is\nrefused. Every upload waits for you to approve the file and the site first. Rejecting it cancels the run, as\nrejecting any gate does, and the file is never attached.\u003C\u002Fp>\n\u003Ch3 id=\"what-the-model-reads\">What the model reads\u003C\u002Fh3>\n\u003Cp>Anything typed into a field that looks like it holds a secret (a password, a token, a key, a PIN), and every cookie\nthe browser holds, is taken out of everything the member reads back: results, addresses, and error messages alike.\nA screenshot is stored with the run and never put in front of the model: the member reads a page with \u003Ccode>snapshot\u003C\u002Fcode>\nor \u003Ccode>extract\u003C\u002Fcode>.\u003C\u002Fp>\n\u003Ch3 id=\"which-models-can-drive-it\">Which models can drive it\u003C\u002Fh3>\n\u003Cp>A run whose task names a browser action starts only on a model trusted to operate a browser. Today those are:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Anthropic:\u003C\u002Fstrong> Claude Opus, Sonnet, Haiku and Fable from version 4 on, and Claude 3.5 Sonnet.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>OpenAI:\u003C\u002Fstrong> 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\nvariant.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>xAI:\u003C\u002Fstrong> Grok from 4.7 on.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Each counts on its own provider or through OpenRouter. On any other model the run stops before the model is called, with\n\u003Ccode>computer_use_model_unsupported\u003C\u002Fcode>, and shows on \u003Cstrong>Blockers\u003C\u002Fstrong>: change the member's model, then retry it. See\n\u003Ca href=\"\u002Fdocs\u002Fcontrol\u002Fspend\">Spend and models\u003C\u002Fa>.\u003C\u002Fp>\n\u003Ch2 id=\"who-connects-it\">Who connects it\u003C\u002Fh2>\n\u003Cp>A browser binding holds no credential. It is three things: a prefix, the sites, and a start URL on one of those\nsites. What the sites see is a signed-out visitor, until a person signs in for the member.\u003C\u002Fp>\n\u003Cp>A binding can sit on the enterprise, a team, a member, a role or a task, and the one further down replaces the one\nabove: task, then role, then member, then team, then enterprise. A binding lower down may leave out sites that a\nwider one it inherits from allows. Adding a site that the wider one does not allow is refused\n(\u003Ccode>widens_&lt;layer&gt;_grant\u003C\u002Fcode>), unless you are the person who bound the wider one, or the CEO.\u003C\u002Fp>\n\u003Ch3 id=\"signing-in\">Signing in\u003C\u002Fh3>\n\u003Cp>A member never types a username or a password itself. When its browser reaches a sign-in page, it calls\n\u003Ccode>crm.sign_in\u003C\u002Fcode>, which its task has to list like any other action, and the run stops.\u003C\u002Fp>\n\u003Cp>The run's page then shows \u003Cstrong>Waiting on a sign-in\u003C\u002Fstrong>, with the site and the page the browser is on. Choose \u003Cstrong>Take the\nbrowser\u003C\u002Fstrong>, sign in by naming a field or a button and typing or clicking, then choose \u003Cstrong>Hand back\u003C\u002Fstrong>. What you type is\nsent to the page only: the run's record and its model never see it. Only someone who may decide the run's gates can\nhand the browser back. The run carries on from the page you left it on.\u003C\u002Fp>\n\u003Cp>What the sign-in left behind (the site's cookies and stored session, never a password) is kept for that member on\nthat site, so the member's later runs start signed in. It is never shared with another member, and a copied task\ndoes not carry it. When the site has ended that session, the next run meets the sign-in page and asks again.\u003C\u002Fp>\n\u003Cp>Each browser binding on \u003Cstrong>Tools\u003C\u002Fstrong> lists the sign-ins it keeps, by member and site, with \u003Cstrong>Revoke sign-in\u003C\u002Fstrong> beside\neach: the next run for that member on that site meets the sign-in page again.\u003C\u002Fp>\n\u003Ch2 id=\"what-waits-for-you\">What waits for you\u003C\u002Fh2>\n\u003Cp>Nothing a browser does waits for you unless the task declares a gate on it, with two exceptions: \u003Ccode>crm.upload\u003C\u002Fcode>\nalways waits for you to approve the file, and \u003Ccode>crm.sign_in\u003C\u002Fcode> waits for a person to sign in.\u003C\u002Fp>\n\u003Cp>Gate one action, such as \u003Ccode>crm.click\u003C\u002Fcode>, and every click waits; gate the prefix, \u003Ccode>crm\u003C\u002Fcode>, and every action does, reading\nincluded. A gated action waits with exactly what it was about to do (the selector, the text, the address), and\nnothing happens on the site until you approve it on \u003Cstrong>Gates\u003C\u002Fstrong>. Approving runs the action together with whatever it\nsends, and no second approval follows. A gate on one action holds no other: a gate on \u003Ccode>crm.click\u003C\u002Fcode> does not hold a\n\u003Ccode>crm.press\u003C\u002Fcode> of Enter that submits the same form. See \u003Ca href=\"\u002Fdocs\u002Fcontrol\u002Fgates\">Gates and tool scopes\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>An approval waits until someone decides it or its run is cancelled. A sign-in is different: one that nobody\ncompletes before your enterprise's gate timeout cancels the run.\u003C\u002Fp>\n\u003Ch3 id=\"what-a-run-does-without-asking\">What a run does without asking\u003C\u002Fh3>\n\u003Cp>On a site its binding allows, a run submits, pays, sends and deletes without asking anyone, unless the task gates\nthe action. What an action sets off goes to the site as a browser sends it: a form that posts, a save in the\nbackground. So does what the page sends on its own between actions. None of it is held or approved separately.\u003C\u002Fp>\n\u003Cp>If you want to see a submission before it happens, gate every action that can submit it: \u003Ccode>click\u003C\u002Fcode>, \u003Ccode>type\u003C\u002Fcode>, \u003Ccode>press\u003C\u002Fcode>\nand \u003Ccode>select\u003C\u002Fcode>, or the prefix. Or keep the binding's sites to pages where nothing can be saved, sent or deleted.\u003C\u002Fp>\n\u003Ch2 id=\"setup\">Setup\u003C\u002Fh2>\n\u003Col>\n\u003Cli>In the portal, on \u003Cstrong>Tools\u003C\u002Fstrong> (or a member's or team's page → \u003Cstrong>Tools\u003C\u002Fstrong>), choose \u003Cstrong>Add tool\u003C\u002Fstrong>.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Kind:\u003C\u002Fstrong> Computer-Use. \u003Cstrong>Tool type:\u003C\u002Fstrong> the prefix your tasks will use, such as \u003Ccode>crm\u003C\u002Fcode>.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Allowed sites:\u003C\u002Fstrong> one host per line, with no \u003Ccode>https:\u002F\u002F\u003C\u002Fcode> and no path: \u003Ccode>crm.example.com\u003C\u002Fcode>. A host covers its\nsubdomains. List only what the work needs.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Start URL:\u003C\u002Fstrong> where the browser opens, on one of those sites: \u003Ccode>https:\u002F\u002Fcrm.example.com\u002F\u003C\u002Fcode>.\u003C\u002Fli>\n\u003Cli>Choose \u003Cstrong>Bind browser\u003C\u002Fstrong>.\u003C\u002Fli>\n\u003Cli>On the task, name each action it needs (\u003Ccode>crm.navigate\u003C\u002Fcode>, \u003Ccode>crm.snapshot\u003C\u002Fcode>, \u003Ccode>crm.click\u003C\u002Fcode>, …), with \u003Ccode>crm.sign_in\u003C\u002Fcode>\nwhen the site needs a sign-in, and declare a gate on each one that should wait for you.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Check it.\u003C\u002Fstrong> Run the task. The run page shows every action, and the address the browser was on after it.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2 id=\"troubleshooting\">Troubleshooting\u003C\u002Fh2>\n\u003Cdiv class=\"prose__table\">\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>What you see\u003C\u002Fth>\n\u003Cth>Why\u003C\u002Fth>\n\u003Cth>What to do\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_not_on_task\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>The task lists the bare prefix, \u003Ccode>crm\u003C\u002Fcode>, not its actions\u003C\u002Ftd>\n\u003Ctd>Name each action: \u003Ccode>crm.navigate\u003C\u002Fcode>, \u003Ccode>crm.click\u003C\u002Fcode>, …\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_site_blocked: &lt;host&gt;\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>The page is on a host the binding does not allow\u003C\u002Ftd>\n\u003Ctd>Add the host to the binding's sites, if the work needs it\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_model_unsupported\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>The run's model cannot operate a browser\u003C\u002Ftd>\n\u003Ctd>Change the member's model to one that can, then retry the run\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_browser_unavailable\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>No browser could be started for the run\u003C\u002Ftd>\n\u003Ctd>Nothing about your task is wrong. Retry the run from \u003Cstrong>Blockers\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Calls fail as though nothing were bound\u003C\u002Ftd>\n\u003Ctd>The start URL is not on one of the binding's sites, so it gives no browser\u003C\u002Ftd>\n\u003Ctd>Put the start URL's host in the sites\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Saving a binding is refused when it adds a site (\u003Ccode>widens_&lt;layer&gt;_grant\u003C\u002Fcode> over the API)\u003C\u002Ftd>\n\u003Ctd>A wider binding it inherits from does not allow that site\u003C\u002Ftd>\n\u003Ctd>Ask whoever bound the wider one to add it there, or leave the site out\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_stale_ref\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>The reference is from an earlier snapshot, or the page has changed since\u003C\u002Ftd>\n\u003Ctd>Nothing was done. The member takes a new snapshot and uses a reference from it\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_snapshot_too_large\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>The page is too large to read in one snapshot\u003C\u002Ftd>\n\u003Ctd>Have the task snapshot one part of the page, fewer levels deep, or only what changed\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_wait_timeout\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>The page did not reach the state a \u003Ccode>wait\u003C\u002Fcode> asked for in time\u003C\u002Ftd>\n\u003Ctd>The browser stays open and the run carries on. Check that the task waits for something the page really shows\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_unknown_option\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>A \u003Ccode>select\u003C\u002Fcode> named a label the dropdown does not offer\u003C\u002Ftd>\n\u003Ctd>The error lists the labels it does offer. The label has to match exactly\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_tab_not_found\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>A \u003Ccode>switch_tab\u003C\u002Fcode> matched no open tab\u003C\u002Ftd>\n\u003Ctd>The error lists the open tabs, by title and host\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_upload_refused\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>An \u003Ccode>upload\u003C\u002Fcode> named something that is not a file the run holds\u003C\u002Ftd>\n\u003Ctd>Have the task upload a file it downloaded, or one the run was started with\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Ccode>computer_use_sign_in_not_completed\u003C\u002Fcode>\u003C\u002Ftd>\n\u003Ctd>Nobody signed in before the enterprise's gate timeout, so the run was cancelled\u003C\u002Ftd>\n\u003Ctd>Start the run again, and sign in from the run's page when it asks\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>A click or a field is not found\u003C\u002Ftd>\n\u003Ctd>The selector does not match the page as it is\u003C\u002Ftd>\n\u003Ctd>Have the task \u003Ccode>snapshot\u003C\u002Fcode> the page to see what is there, and act by reference\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n",[13,17,21,24,27,30,33,36,39,42,45],{"id":14,"text":15,"level":16,"planned":9},"what-a-member-can-do","What a member can do",2,{"id":18,"text":19,"level":20,"planned":9},"only-the-sites-you-allow","Only the sites you allow",3,{"id":22,"text":23,"level":20,"planned":9},"files-in-and-out","Files in and out",{"id":25,"text":26,"level":20,"planned":9},"what-the-model-reads","What the model reads",{"id":28,"text":29,"level":20,"planned":9},"which-models-can-drive-it","Which models can drive it",{"id":31,"text":32,"level":16,"planned":9},"who-connects-it","Who connects it",{"id":34,"text":35,"level":20,"planned":9},"signing-in","Signing in",{"id":37,"text":38,"level":16,"planned":9},"what-waits-for-you","What waits for you",{"id":40,"text":41,"level":20,"planned":9},"what-a-run-does-without-asking","What a run does without asking",{"id":43,"text":44,"level":16,"planned":9},"setup","Setup",{"id":46,"text":47,"level":16,"planned":9},"troubleshooting","Troubleshooting",1794928143303]