Human checks

Use a human check when the page needs a person. Swarmtab does not solve CAPTCHAs by program. An agent can call:

{"name":"ask_human","arguments":{"message":"Please complete the check in this page, then return control."}}

The agent enters needs_you. Open the request in the Swarmtab window, inspect the page, and take control. Complete the check by hand. Use the window’s completion control to resume or decline it. The agent can repeat the identical request to wait.

While you have control

The agent is locked out of page reads and actions, including screenshots. It receives: “the human has this page now; call ask_human again to wait”. After you release control, its next read starts fresh.

Not every website challenge can be detected automatically. If a task stalls at a check, ask the agent to call ask_human or take control from the window. Never ask it to bypass the challenge.

Enter a code or secret

For supported secure-entry controls, the agent can request ask_human with kind: "code" or kind: "secret" and the control ref. Secure entry blocks scripts and exports and blacks out page pictures until the entry closes. Only enter secrets into the intended site after checking its address.

Permission approvals are separate from page checks. Read the action and scope before approving either. A page’s request does not grant it authority over Swarmtab.