AI Agents

Giving AI agents phone intelligence tools

An autonomous fraud triage or KYC agent that reasons over a phone number still needs a real, live signal to reason with, not a static description of what a phone number usually means. Exposing Telebase's lookup as a callable tool lets an agent decide for itself when a carrier, number type or active-status check is worth making, rather than a human wiring that decision into fixed business logic ahead of time.

Why agents need tools, not just training data

A language model has broad knowledge about what a VoIP number is or how a SIM swap enables fraud, but it has no way to know whether a specific number in front of it, right now, is active, which carrier serves it, or whether it is a VoIP line. That is live, per-query network state, and the only way an agent gets it is by calling something that queries the network. This is the same gap that made web search and code execution necessary agent tools: the model reasons, the tool supplies ground truth.

What a phone intelligence tool call looks like

Structurally, a Telebase tool definition is a thin wrapper around one HTTP call: the agent framework describes the function (name, that it takes a phone number, what it returns), the model decides when to invoke it, and the response comes back as structured JSON the model can reason over directly.

{
  "name": "lookup_phone_number",
  "description": "Look up carrier, country, number type and active status for a phone number in E.164 format.",
  "parameters": {
    "phone": "string, E.164 format, e.g. +447700900000"
  }
}

The agent's own reasoning loop handles when to call it: before releasing funds on an unfamiliar number, before completing onboarding for a new account, or when a conversation surfaces a phone number that needs verifying. The tool does one job well and returns a predictable shape every time, which is what makes it safe to hand to a model rather than a person.

Machine-readable documentation matters as much as the API itself

An agent building its own tool call, or evaluating whether Telebase fits a task, benefits from documentation written for a model to parse directly rather than a human to read in a browser. Telebase publishes a machine-readable API guide at skills.md for exactly this reason: exact request shape, exact response fields, and no marketing language for an agent to have to filter out.

What this does not change

The live vs launching distinction still applies regardless of who, or what, is calling the API. An agent querying a GB, DE, NL or French number today gets simSwap back as UNKNOWN, the true value while carrier registration completes, the same as a human-triggered call would. An agent should be built to treat UNKNOWN as "no signal available", not silently interpret it as "no swap detected".