Phone calls for Vercel AI SDK agents: the call4me tools package

call4me-ai-sdk gives an AI SDK agent a phone. It calls a business, has the conversation, waits on hold, and hands the outcome and transcript back to your code. Here is the setup, the tools, how mid-call questions work, and what happened when we ran it.

Nick Khami · October 8, 2026 · 4 min read

The short answer: install call4me-ai-sdk, set CALL4ME_API_KEY, and pass call4meTools() to generateText or streamText. Your model gets a makePhoneCall tool that calls a business, has a natural conversation (booking, rescheduling, cancelling, asking, waiting on hold, getting through phone menus), and returns the outcome and transcript when the call ends.

npm install call4me-ai-sdk ai zod @ai-sdk/openai

import { generateText, isStepCount } from 'ai';
import { openai } from '@ai-sdk/openai';
import { call4meTools } from 'call4me-ai-sdk';

const { text } = await generateText({
  model: openai('gpt-5-mini'),
  prompt:
    'Call Pottery Barn customer service at 1-888-779-5176 and ask how long furniture delivery ' +
    'usually takes and whether returns have a shipping fee. General questions only, no order.',
  tools: call4meTools(),
  stopWhen: isStepCount(8),
});

console.log(text);

The package has no runtime dependencies of its own. Every tool is a thin call to the call4me MCP server, so an AI SDK agent and a Claude Code or Codex session get the same caller, the same phone numbers and the same account.

Get an API key #

1. Sign in at call4.me and add credits. Calls cost $0.25 per minute of talk time, held up front and settled when the call ends. Unanswered calls are free.
2. Copy your key (it starts with cb_live_) from your account page.
3. Set it as CALL4ME_API_KEY, or pass it directly: call4meTools({ apiKey }).

Keep the key on your server. It can place calls billed to your balance.

The tools #

call4meTools() returns an object you can pass straight to tools, or pick from:

Every tool returns { text, isError, data }: text is a readable summary the model can use directly, and data is the structured result (call id, status, outcome, open questions) when the server sends one.

Options:

call4meTools({
  apiKey: process.env.CALL4ME_API_KEY, // default
  maxWaitMinutes: 15, // how long makePhoneCall waits before handing back the call id
});

How a call works inside one generateText run #

Phone calls take minutes, not milliseconds. makePhoneCall places the call, then follows it with getCall (each request waits up to 50 seconds on the server) until one of three things happens:

1. The call finishes. The tool returns the outcome (done, partial, not_possible, voicemail and so on), a summary, structured details and the transcript.
2. The business asks something only the user knows (a date of birth, "does 8:15 work?"). The tool returns early with open_questions. The model answers with answerQuestion, then follows the call with getCall. The business is waiting on the line, so answer quickly.
3. maxWaitMinutes passes. The tool returns the call's current state and id, and the model can keep following it with getCall.

Give the run enough steps for that loop. isStepCount(8) covers a call plus a question or two.

Before it dials, call4me checks that it knows what the call needs. A restaurant booking needs a party size, a date and a time window. A general question needs the questions. If something is missing it does not dial: the tool result says exactly what to ask the user, and the model can ask and try again. getRequirements lists the fields for each kind of call.

Following a call yourself #

For a UI that shows the call as it happens, or a job queue, use placeCall and getCall instead of makePhoneCall:

import { createCall4meClient } from 'call4me-ai-sdk';

const call4me = createCall4meClient(); // reads CALL4ME_API_KEY

const placed = await call4me('call4me_place_call', {
  to: '+18887795176',
  business: 'Pottery Barn customer service',
  category: 'general',
  goal: 'Ask how long furniture delivery usually takes and whether returns have a shipping fee.',
  details: { questions: 'How long does furniture delivery usually take? Is there a return shipping fee?' },
});
const callId = placed.data?.id as string;

const view = await call4me('call4me_get_call', { call_id: callId, wait_seconds: 50 });
console.log(view.data?.status, view.data?.transcript);

Repeat the call4me_get_call request until data.finished is true, rendering data.transcript as it grows and answering anything in data.open_questions with call4me_answer_question.

What we tested #

We ran the package on October 8, 2026 with ai 7.0.135, @ai-sdk/openai 4.0.91, gpt-5-mini and call4me-ai-sdk 0.1.0, on our own account.

Balance check. Asked "How much call4me credit do I have left, and what number do my calls come from?", the model called getBalance and answered from the live result: the balance and the account's three numbers.

A real call. We asked it to call Hulu support at 877-824-4858 and ask whether Hulu has a student discount and how a student proves eligibility. The model called makePhoneCall, which placed the call and waited. A live representative answered and asked the question back, then the call ended after about 3 minutes: Hulu's side hung up before answering. The tool returned the outcome as partial with the transcript, and the model told us the question was not answered and offered to call again. That is the behavior you want from a tool: it reports what actually happened on the call, not what was supposed to happen.

Pricing and limits #

• $0.25 per minute of talk time in the US and Canada; some destinations cost more (see the FAQ).
• Two calls can run at once per account.
• Set max_minutes in a call's input to cap its talk time and leave credits free for other calls.

Use it from Claude Code or Codex instead #

The same calls work from coding agents without any code: see Claude Code phone calls and Codex phone calls, or the MCP setup page.

The package source is MIT licensed, in the call4me repository.

vercel ai sdk · ai sdk tools · phone calls · agents · mcp