RaabtaHQ

Back office, overnight

Reaching into a system we have never heard of

Your own API, as a tool the assistant can use

Every business has one system nobody else has — an inventory database, a warehouse API, a legacy booking tool. The assistant being unable to see it is usually the reason automation stops at the interesting part.

  • Describe your endpoint once; it becomes a tool
  • Same approval gates as any built-in capability
  • Outbound calls pass a guard — HTTPS only, no internal addresses
  • Credentials encrypted, and never shown to the model

Today

An assistant that can answer everything except the question customers actually ask, because the answer lives elsewhere.

With this running

Your own system answering through the same conversation, with the same approvals and the same audit trail.

Step by step

What actually happens

  1. Your team

    Describes the endpoint — method, URL, fields, credentials — once, in settings.

  2. Customer

    “Do you have the blue one in stock in Lahore?”

  3. Assistant

    Calls your inventory endpoint like any other tool and answers from the response.

  4. Your team

    Sees the call, its arguments and its result on the run's timeline, like everything else.

01 / 02

It is a tool, not an exception

A connector goes through the same machinery as a built-in capability: the per-assistant permission map, the approval gate, the audit trail, the spend ceiling. There is no separate path with weaker rules — which is what makes it safe to add one without re-reasoning about the whole system.

02 / 02

The guard on the way out

Outbound calls are forced through a guard: HTTPS only, private and internal addresses blocked after the hostname is resolved, and redirects refused. A connector cannot be pointed at something inside your own network, deliberately — that is the failure mode that turns a helpful integration into a way in.

  • HTTPS enforced, plaintext refused
  • Resolved addresses checked, so a public name cannot map inward
  • Redirects not followed
  • Credentials encrypted at rest and never in the model's context

How this is put together

The same building blocks every assistant is made of. Nothing here is bespoke to this scenario.

What starts it

A customer messageAny run — a connector is just another tool

Where it runs

WhatsAppFacebook MessengerInstagram DMWebsite chat

What it can do

  • Looked up a contact by number
  • Looked up the contact
  • Replied
  • Added a note

What it uses

A connector to your own systemYour instructions library

Every action above is a real capability, and every run keeps a record of which ones it used. How the engine holds this together →

More like this

Other things it does

Browse the whole library

Would this work for how you actually operate?

Tell us the scenario you have in mind. If it needs a building block we do not have, we would rather say so than sell you a demo.