Page Agent

Experimental — developer preview. Page Agent is opt-in and off by default. It is an early feature and its behaviour, scope, and controls may change. Enable it only if you’re comfortable with a preview.

Page Agent is an in-page agent that operates the HivemindOS dashboard for you from plain language. You type what you want — “open the work board and add a task”, “switch to the agents view” — and it reads the current screen, decides which control to use, and clicks and types for you.

It is built on the open-source alibaba/page-agent project (MIT), driven through your own configured model.

How to try it

Page Agent currently lives in the developer lab; it is not an installable My Apps card and it does not mount across the normal dashboard. In a development build, open /page-agent-lab. Production builds keep that lab unavailable unless the operator starts the dashboard with PAGE_AGENT_LAB=1.

The lab shows a command bar. Type an instruction and Page Agent drives only that lab page.

What it can — and can’t — reach

Page Agent works inside the HivemindOS window only. It reads and controls the dashboard you’re looking at.

• It cannot reach out to a separate Chrome, Safari, or other browser window, and it cannot control other desktop apps. Those need a different tool (see the browser-automation and computer-use capabilities), not Page Agent.
• It does not take screenshots or use your camera. It reads the page’s structure as text and acts on it.

Safety

The dashboard can move money (wallets, trading), so Page Agent is deliberately constrained:

• Money is off-limits. Page Agent is kept away from wallet, trade, and transfer controls — moving funds always goes through the normal confirm-first wallet flow, never an automated click.
• No arbitrary code. The tool that would let it run generated JavaScript is disabled and stays disabled.
• Fresh state before interaction. Click, type, and selection tools re-read the page immediately before acting instead of trusting an old element number.
• Untrusted pages pause the run. Instruction-like page content is treated as data and suspected prompt injection stops the task for human review.
• Consequences ask at the moment of action. Submit, send, publish, purchase, transfer, delete, install, upload, and signature-shaped controls require a one-action confirmation.
• Receipts survive reloads. Governed runs retain redacted observations, policy decisions, action results, and post-action verification without storing typed text or secrets.
• You stay in control. It’s opt-in, off by default, step-capped, and can be stopped mid-task at any time.

Limitations

• Preview quality: it can misread an unusual control or take an extra step or two and correct itself.
• It needs a model that supports tool calling; if your active agent’s model can’t, Page Agent will tell you rather than guess.
• It operates one screen at a time — it doesn’t span multiple windows or tabs.
• The developer lab is not a promise that every dashboard control is ready for autonomous operation. For general web browsing, HivemindOS uses the governed computer-interaction capability and Browser Use instead.