Skip to content
Agentic Workflowdocs
v0.9.0Install free

nodes / builtin / ai / aidependencies / tools

Browser

Give an agent a browser: look at, read and use web pages, within the sites and actions you allow.

AI ExtensionWeb app — not availableAgent toolSince 0.9.0

Give an AI agent a browser. The Browser block lets an agent open pages, look at them and read them, and, when you allow it, click, type and pick options. One block, your rules.

Browser is a building block for an agent, not a step in the flow. Connect it to the Tools slot of a Tools Agent (from the agent’s Building blocks, + Add a tool). The agent then uses the same browser engine as Aria: Look returns the page’s elements, each with a short reference, and actions name an element by its reference, not by a CSS selector.

What the agent may do depends on What it can do:

  • Read only (default): open a page, look at it, scroll and read. Nothing on the page changes.
  • Read + act: also click, type and pick options. Choose which in Act (all three when empty).

The rules that keep it safe:

  • Allowed sites: the agent opens and acts only on these sites. Opening a new site asks you first; Always allow adds it to the list. Acting on a site that isn’t allowed is refused.
  • Acting after reading asks: once the agent has read a page in this run, an action asks you first, even on an allowed site. Page text could contain instructions meant to hijack the agent.
  • Submitting always asks: a click that submits a form, or Enter in a form, always asks.
  • Page content reaches the model fenced as data, so instructions inside a page are not followed as commands.
  • Ask before a call sets when the agent asks you for the other actions (see Action approval).

Browser works in the extension only. In the web app, the agent is told the browser isn’t available and answers without it.

  • An agent that researches: open a few pages, read them, and summarise.
  • An agent that works through a page for you (fill a form, pick options), with you confirming the risky moments.

For a fixed, planned sequence of page actions, use the regular page steps instead: they need no model and no approval.

SettingTypeWhat it does
What it can do ChoiceOptions: Read only, Read + act. Default: Read only
Act Multiple choiceWhat it may do on a page (with Read + act). Options: click, type, select.
Ask before a call ChoiceWhen the agent must ask you before this tool runs. Options: Every call, Only calls that change things, Skip trusted sites, Ask me anywhere, even when I’m away.
Allowed sites TagsThe agent opens and acts only on these. “Always allow” on a new site adds it here.

Allowed sites is cleared when you publish the workflow: whoever installs it picks their own sites.

This node has no output port. It plugs into a slot on another node instead of passing items along.

What each action did goes back to the model, not to the next step. The agent’s own output shows the result: its response, and the tools it called in steps.

None. This node needs no connection and has no slots for other nodes.

  1. Keyboard shortcut starts the workflow on the page you are reading.
  2. Tools Agent with “Find the pricing page of this site and list the plans”, a chat model, and Browser (Read only, Allowed sites set to the site) in Tools.
  3. Show data shows {{ $input.response }}.
  • “Action denied: … is not one of the allowed sites.”: add the site to Allowed sites, or approve it with Always allow when the agent asks to open it.
  • The agent waits: an action is asking for your approval. Approve or reject it in the run.
  • “The browser is not available here”: the workflow ran in the web app. Run it from the extension.

Workflows made before AWFlow 0.9 may use Browser Tool, which acts on the active tab with CSS selectors (click, fill, type, select, submit, navigate) and asks according to its Ask before a call setting. It keeps working. For new agents, use the Browser block.

Edit page

Last updated:

Was this page useful?
Ask Aria about Browser