Agents
App Access
App access lets an agent work with the page a user is on: read its contents, point out a control, fill in a form, or navigate to another part of your app. These tools run through the embedded assistant in the user's browser. They are separate from the API operations you register under Actions.
Open the agent's App access settings to enable each tool. For an enabled tool, Allow runs it without an approval request; Ask pauses for approval before it runs. Changes to these settings take effect for users when you publish the agent.
Read the current page
Read the current page gives the agent a structured view of the page's text, controls, and field values. It can use that information to understand what the user is looking at and identify elements to highlight or fields to fill.
You can exclude parts of your app using Ignored regions (CSS selectors). Enter one selector per line. Matching elements and their contents are left out of page reads. You can also mark an element directly in your frontend:
<section data-teaflask-ignore>
{/* Content excluded from page reads */}
</section>
Password-input values are redacted. The reader also redacts input and textarea values marked with sensitive autocomplete values, including passwords, card numbers, card security codes, card expiry dates, and one-time codes. For other content you want excluded, use ignored regions.
Navigate the app
Navigate the app lets the agent take the user to an app-relative path, such as /settings. Configure Allowed path prefixes with one prefix per line:
/allows navigation throughout the app./settingsallows/settingsand paths beneath it, such as/settings/profile, but not/settings-extra.
Enabling navigation initially sets the prefix to /. Replace it with narrower prefixes if the agent should navigate only within particular sections.
Your frontend supplies an onNavigate callback that connects the assistant to your app's router. See Install for wiring it in. Without that callback, the assistant cannot perform navigation.
Navigation prefixes restrict where the agent may take a user; ignored regions separately control what page reads include. Adding a page to the sitemap does not grant navigation permission.
Highlight on the page
Highlight on the page scrolls to an element and spotlights it so the user can find it. It does not activate the highlighted control.
Enable Read the current page alongside highlighting. The agent uses the element references returned by a page read to identify what to highlight, rather than guessing a selector from the user's description.
Prefill forms
Prefill forms lets the agent enter values into form fields for the user. Enable Read the current page alongside it so the agent can identify the fields.
Prefilling supports text inputs, textareas, native dropdowns, checkboxes, radio buttons, and editable text regions. Password fields and input or textarea fields marked with the sensitive autocomplete values described above are refused. File pickers, disabled fields, and read-only text fields cannot be filled.
The tool fills fields; it does not press a submit button. Your app's normal input and change handlers still run, including any autosave behavior your app implements.
The agent receives a result for each requested field, including fields that could not be filled. If only part of a form can be filled, successful edits remain in place; they are not rolled back because another field failed.
Frontend requirements
Page reading, highlighting, and form prefilling are built into the embedded assistant. They do not require an actions adapter or a separate handler for each control. Navigation requires your onNavigate callback.
To let the agent perform registered API operations as well, configure Actions and an adapter. That is a separate capability from interacting with the page.