HivemindOS manual

Page Agent

Experimental developer preview. Page Agent is off by default, and its behavior and controls may change.

Page Agent operates the HivemindOS dashboard for you from plain language. Type something like “open the work board and add a task”, and it reads the screen, picks the right control, and clicks and types for you. It is built on the open-source alibaba/page-agent project (MIT) and uses your own configured model.

Try it

Page Agent lives in a developer lab page. It is not a My Apps card and doesn’t run across the normal dashboard.

In a development build, open /page-agent-lab. In a production build, the lab is available only if the dashboard was started with PAGE_AGENT_LAB=1.
Type an instruction in the command bar. Page Agent acts only on that lab page.

What it can reach

Only the HivemindOS window. It can’t control a separate browser window or other desktop apps. For those, HivemindOS uses Computer Interaction and Browser Use.
It reads the page structure as text. It doesn’t take screenshots or use your camera.

Safety

Money is off-limits. Page Agent can’t use wallet, trade, or transfer controls. Moving funds always goes through the normal confirm-first wallet flow.
No generated code. Its ability to run generated JavaScript is turned off.
Fresh state. It re-reads the page right before each click or keystroke.
Untrusted pages pause the run. Instruction-like page text is treated as data, and suspected prompt injection stops the task for your review.
Consequences ask first. Submit, send, publish, purchase, transfer, delete, install, upload, and signature-like controls need a one-time confirmation.
Receipts. Runs keep redacted records of what it saw, decided, and did, without storing typed text or secrets.
You stay in control. It’s opt-in, limited in steps, and you can stop it at any time.

Limitations

As a preview, it can misread an unusual control or take an extra step to correct itself.
It needs a model that supports tool calling. If yours can’t, Page Agent tells you instead of guessing.
It works on one screen at a time, not across windows or tabs.
Not every dashboard control is ready for autonomous use.