Approve an agent’s action before it happens

Approve the exact action, not a vague intent

A useful approval names precisely what it allows. “The agent may send the invoice” is too broad; “send invoice 1042 to Acme for $4,280” is what a reviewer can actually check. Toolcaise approvals are bound to the exact action your code is about to perform.

1. Your code describes the action as JSON: every parameter that matters for execution.
2. The SDK sends Toolcaise a safe title and a digest of that JSON, never the whole action: only the fields you choose to show, redacted.
3. A person approves or denies it in the dashboard, seeing the title and those fields as reported by the agent. If the agent’s last decided request in the past 30 days with the same title and destination had other parameters, the dashboard says so.
4. Your code receives a frozen copy of the approved action and executes exactly that. The approval is consumed once and cannot be reused.

Add it to your code

def send(action):
    return billing.send_invoice(action["invoice"], action["to"])

client.guarded_action(
    "Send invoice 1042 to Acme",
    {"invoice": 1042, "to": "acme", "amount_usd": 4280},
    send,
    show_fields=["invoice", "to", "amount_usd"],
)

If the person denies it, the approval expires, or Toolcaise cannot be reached to confirm the decision, the call raises and your function never runs. Approvals fail closed in every mode.

Pause, resume and cancel

Beyond single actions, a run can be paused, resumed or canceled from the dashboard. Through the SDKs these are cooperative: your code calls checkpoint() between safe steps, where it waits while paused and stops cleanly when canceled. For a stop that does not depend on the agent’s cooperation, run it under Toolcaise Connect, which can end the whole process tree.

Rules for MCP tool calls

If your agent works through MCP tools, put Toolcaise Connect’s gateway in front of the MCP server to allow or deny tools by name and cap how many calls a session makes. With --remote-control, the dashboard can also change those rules and hold a specific tool call for a person’s approval.

Who can approve, and what gets recorded

Workspace owners and admins decide approvals. Each decision records who made it, when and any note they left, and the workspace shows the approver’s email address.

• An approval that nobody decides expires. It never quietly turns into a yes.
• If the run that asked for it ends, a pending approval lapses and can no longer be granted.
• A denial stays a denial; it cannot be reversed into an approval for the same action.

Where approvals stop

An approval proves a person allowed an action. It does not prove the action happened or that the destination received it: the external system is the only source of that. Approvals also only guard code that goes through the integration, so an action your agent performs another way is not covered.