Skip to main content
A tool with approval pauses the reply before execute runs. The call is saved with the chat as waiting for approval, the turn stops as awaiting, and the reply continues in the same assistant message once someone answers. Other calls in the same step still run.

In the chat

The admin chat docks the waiting call above the composer, which stays hidden until every waiting call is answered. The card shows the tool, its input, the rule’s reason when there is one, and Approve and Deny. Deny opens an optional reason for the agent. An approved call runs the input the model sent, as stored with the chat. An answer can’t change it.

Auto-approve

When an agent has a tool, or an MCP server tools entry, that sets approval, the composer shows an Auto-approve switch.
  • While it’s on, 'ask' calls run without asking. 'required' calls still pause.
  • The switch is saved on the chat as its autoApprove checkbox, so it applies to every later turn and to a turn resumed after an answer.
  • Turning it on doesn’t answer calls that are already waiting.
The chats collection’s update access decides who can change it. On a new chat, the switch is sent as autoApprove with the first message and saved when the chat is created. A run started in code reads the chat’s autoApprove unless it sets its own:

Nobody to ask

agent.generate(), workflows, jobs, the MCP plugin and channels whose piece can’t ask questions can’t pause. There, a call that needs approval is denied and the model reads:
The run keeps going. 'ask' tools run when the run sets autoApprove: true or the chat has it on; 'required' tools are always denied.

Who can answer

Whoever can continue the chat can answer, the same as for questions:
  • Over HTTP, the chat’s owner.
  • In a channel conversation, any participant who passes the agent’s access check.
FrogBot records each answer, who gave it, and when, in the message’s settlements field, with the outcome approved or denied.

Over HTTP

A turn that pauses responds with status: "awaiting-input". A waiting approval has an approval object, with the rule’s reason when it gave one:
Answer it through the settle endpoint:
Errors are the same as for questions:
  • Sending an output for an approval, or approved for a question, is 400 with code: "invalid-output".
  • A second answer is 409 with code: "already-settled".
A chat UI built on the AI SDK can also answer with addToolApprovalResponse and resubmit the assistant message. FrogBot reads only each part’s toolCallId, approval.approved and approval.reason, and settles them against the stored call. A resubmitted denial needs a reason; deny without one through settle.

In channels

A channel piece that supports questions asks for approval with the same controls, so no piece needs changes. The approval is posted as a question with the header Approval, the text Approve <tool>? and the rule’s reason, and the options Approve and Deny:
  • Approve approves the call.
  • Deny denies it with no reason, which ends the turn.
  • A typed answer denies it, with the text as the reason the model reads.