Goal

Build an agent that automates a browser task: navigate to a website, extract data, and save it to a file — all with human-in-the-loop approval for sensitive actions.

Prerequisites

  • An Allternit API key
  • The ACU gateway running locally (http://127.0.0.1:8760)
  • @allternit/sdk installed

Implementation

1. Set up the computer capability

2. Start a browser run

3. Stream events

4. Approve actions programmatically

5. Save the results

Explanation

  • allowedSites — Restricts the browser to specific domains. The agent cannot navigate to other sites.
  • autoVerify: false — Requires human approval before each action. Set to true for read-only tasks.
  • Vision coordinates — The agent uses screenshot-based vision to understand the page layout. Pixel coordinates are mapped to the display dimensions.

Next steps

  • Add multi-tab support for tasks that require navigating between pages.
  • Use ACI in Gizzi Code — the CLI has built-in computer-use support via the computer tool.
  • Combine with the memory store to persist extracted data across runs.