HivemindOS manual
Computer Interaction
Computer Interaction is how an agent operates the HivemindOS dashboard or a web page under your rules. Each job is one run that keeps its observations, decisions, actions, approvals, and checks together, even if you reload or resume later.
HivemindOS picks the most specific way to act:
- A built-in Hive action for a known product task.
- Bee Pilot for typed dashboard actions.
- Page Agent, inside its opt-in developer lab.
- Browser Use for general web pages.
- A fresh screenshot only when nothing more structured works.
A named action such as “open the Work board” is easier to check and explain than a guessed click position.
Safety
- Limits. Each run can list allowed websites and apps, plus maximum steps, time, and cost. If the run can’t confirm it is still on an allowed site or app, it stops.
- Fresh state. Every action must be planned from the latest view of the page.
- Untrusted pages. Page text is treated as data. Suspected prompt injection pauses the run so you can inspect it.
- Approval for consequences. Submitting, sending, publishing, buying, transferring, deleting, uploading, installing, running code, and similar actions stop for an approval tied to that exact action. An old or different approval won’t resume it.
- Receipts. After each action, HivemindOS looks again and records the result. Receipts replace typed text, credentials, and secret-like values with redaction markers.
Pause, resume, and stop
You can pause, resume, or stop a run without losing its limits, latest view, receipts, or history. Completed, failed, and stopped runs stay finished; they never restart on their own.
Browser runs keep their named Browser Use session, so a resumed task continues in the same browser. Browser Use is optional and has its own limited or full permission setting. Uploads and generated JavaScript need full permission.
What this does not do
Computer Interaction is not blanket control of your computer. Page Agent stays inside its lab page, Browser Use stays inside its controlled browser, and a screenshot is evidence, not permission. Wallets, payments, trades, and other protected actions keep their own confirmation flows.