Use case

Let agents act, and keep a person on the decisions that matter

The model is an untrusted planner. It proposes a tool call; Caterfli checks that the tool exists, that the agent may use it, that your workspace owns the connection, that the person holds the permission and that the arguments match the schema. Then anything risky pauses for a human — and the approval authorises exactly that call, once.

In practice

What typically waits for a person

  • “Deleting a customer record”
  • “Sending an email from the built-in toolkit”
  • “An unrecognised action on an MCP server, such as a refund”
  • “A destructive call one agent makes on another’s behalf”

What keeps it safe

  • ✓ High-risk and destructive tools need approval by default
  • ✓ An approval is bound to one call and its exact arguments
  • ✓ A pending approval expires rather than waiting forever
  • ✓ Every decision is recorded in a tamper-evident audit trail
How approvals work

How it works

From connection to answer

  1. 01

    The model proposes

    A tool call and its arguments — nothing more. It cannot run anything itself.

  2. 02

    The platform checks

    Existence, ownership, permission, schema and policy, independently of what the model says.

  3. 03

    A person decides

    Risky calls wait in the approval queue with their exact arguments. Approve, and that one call runs.

FAQ

Questions that come up

Something not covered here? Ask us directly.

Which tool calls need approval?

By default, every tool rated high risk or above — deletes, and actions such as sending email. You can require approval on any other tool, or relax it, per agent.

Can an approval be reused?

No. An approval covers one call with the exact arguments a person reviewed. If the model changes the arguments or re-plans, it has to ask again.

What if nobody responds?

A pending approval has a deadline. When it passes, the request expires, the call never runs, and the run it was holding is recorded as failed.

Does approval work in Slack, Telegram and WhatsApp too?

Yes. Every channel reaches the same runtime and the same approval gate as web chat.

Try it on your own system

Connect one system and read the agent Caterfli writes from it. There is a demo connection too, if you would rather look before you wire anything up.