Let the agent do the work, not just describe it.

Give your voice agent carefully scoped access to your APIs. It can retrieve live information, submit approved actions, and explain the result naturally while the caller is still on the line.

Tool action

Talk To My Agent

LIVE

Caller asks

Can you check whether order 1842 shipped?

Secure tool call

GET /orders/1842 · authenticated · 612 ms

Agent answers

It shipped this morning and arrives tomorrow.

Live data returned

Answers are useful. Completed actions are valuable.

Turn the phone call into a working interface for the systems your business already owns.

Use current data

Look up availability, order status, customer records, prices, balances, or any other fact exposed by an approved API.

Take bounded action

Create a ticket, update a record, start a return, or trigger your own workflow with explicit parameters and server-side validation.

Keep your architecture

Tools live with the self-hosted gateway and call your HTTPS endpoints directly. Your systems remain the source of truth.

A controlled path from conversation to API.

Tool definitions tell the model when to call, what to collect, and how to turn a structured response into a clear spoken answer.

  1. 01

    Define the tool

    Name the action, describe when it is allowed, and provide a strict input schema.

  2. 02

    Keep secrets out of prompts

    Reference server-side environment variables for credentials instead of putting tokens in the tool file.

  3. 03

    Validate at your endpoint

    Authenticate the request, authorize the deployment, validate every field, and reject anything outside policy.

  4. 04

    Return only what should be spoken

    Use a response template so the agent receives the useful result, not an oversized internal payload.

Designed for least privilege on a server you control.

A tool is not a blanket permission. Each capability has its own method, endpoint, schema, timeout, authentication, and public-or-private scope.

Private-only tools

Owner tools can be omitted entirely from public calls, so a customer-facing agent never sees them.

Validated configuration

Reserved names, unsafe URLs, malformed schemas, and unsupported settings fail validation before a call begins.

Fast direct execution

Frequent API actions can run as direct HTTPS tools without a slow round trip through the general agent backend.

What one tool can unlock

Read the custom tools reference
Real-time inventory lookup
Order and account status
Quote calculation
Ticket creation
Return authorization
A custom workflow in your own API

Tool action questions

Can this call any API?

It can call approved HTTPS endpoints that match the tool validator and your network policy. Your endpoint remains responsible for authentication, authorization, and business validation.

Where are API credentials stored?

Credentials stay in environment variables or the appropriate secret store on the server. The tool definition references them; it should not contain raw tokens.

Can public callers use owner tools?

Not when the tool is marked private-only. It is omitted from the model’s available tools on public calls, which is stronger than relying on a prompt instruction.

Do I need to modify the gateway source?

No for ordinary custom HTTPS tools. Add a validated entry to voice-tools.json. Packaged Voicebridge apps use the same tool boundary for deeper integrations.

Your API already knows how the business works.

Expose one narrow action and let the agent complete it naturally on the next call.

Design a tool with us