Give your agent a phone

Hollerback lets an AI agent get things done by phone. Your agent says who to call and what it needs to find out. Our voice agent makes the call, tries again if nobody answers, and hands back a short, structured answer.

How it works

  1. Your agent writes a task.

    A task is a short brief, like one you would give a new assistant: who to call, why, which facts to share, and which answers to bring back.
  2. The voice agent calls.

    It says who it is calling for, checks that the right person answered, and works toward the goal. It shares only the facts in the task. If nobody answers or the person asks for a callback, it calls again later.
  3. Your agent gets the answer.

    Each task ends with a state such as goal_met or declined, a one-line summary, and the fields your agent asked for.

What goes in a task

FieldIn plain wordsExample
personWho to call.Dana Ruiz, +1 415 555 0134, Pacific time
goalWhat the call is for, and what counts as done.Ask Dana to cover Tuesday's shift. Get a yes or no.
contextEvery fact the voice agent may say. It says nothing else.The shift is 9am to 5pm at the Elm Street store.
authorityWhat the voice agent may offer or agree to. Empty means nothing.You may offer a $20 bonus.
resultSchemaThe answers your agent wants back, as fields.accepted: true or false

Your agent does not need to learn these fields from this page. When it connects, Hollerback gives it the same instructions.

Set it up

  1. Get an API key.

    Sign in, add funds, and buy a phone number. Then create a key on the API keys page. The beta approves each account by hand.
  2. Connect your agent.

    Pick the app your agent runs in. Copy the setup and paste it where the instructions say.

    Paste a key to fill it into the setup below. The key stays in this browser tab.

    Run this once in your terminal. Claude Code then has the Hollerback tools in every session.

    claude mcp add --transport http --scope user hollerback https://hollerback.dev/api/mcp \
      --header "Authorization: Bearer <your API key>"
  3. Ask for a call.

    Talk to your agent the way you would talk to a person. For example:

    Call Dana Ruiz at +1 415 555 0134. Ask if she can cover the Tuesday shift at Elm Street, 9 to 5. She agreed to work calls when she was hired. Tell me what she says.

    The agent writes the task, Hollerback makes the call, and the agent tells you the answer. Calls happen between 8 a.m. and 9 p.m. in the person's time zone, so the answer can take a while.

When the person asks something new

The voice agent never guesses. If the person asks a question the task does not answer, the voice agent says it does not know and that someone will get back to them. If the goal cannot move forward without that answer, the task waits in waiting_for_answer.

Your agent reads the question, finds the answer, and sends it with answer_question. The voice agent calls the person back with the answer. To avoid the extra call, put likely questions and their answers in context.

Show options on the person's phone

Some answers are easier to pick than to say, such as a room type or an appointment time. Give the task a list of companion options. During the call, the voice agent asks the person if it may text a link. The link opens a page on their phone with your options. The person taps one and confirms. The voice agent hears the choice at once, and your agent gets it as a selection on the task.

Each option has an id, a title, and optionally a description, a short detail such as a price, and an image. The voice agent writes only the heading of each screen. The text and images on the cards come from your options.

Set it up

  1. Register for texting on the Texting page. The link goes out under the same registration as your other texts, with no extra step or fee. A number can send links once its companion state in get_texting_status is on.
  2. Create the task with companion.options, a call channel, and consent.texts set to true.
  3. Read companion on the finished task. It lists every screen shown and every selection.

What the person sees

  • The text says who is writing and that it is an AI assistant, and it asks them to reply STOP to opt out.
  • The page shows your company name and the current options. It never shows their phone number or the transcript.
  • The page stays open for 30 minutes after the call ends, then the link expires.
  • The link text costs the same as any other text. The page is free.

Three ways to connect

All three use the same API key and do the same things. Pick the one that fits where your agent runs.

MCP server

Your agent runs in Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, Gemini CLI, or the Claude or OpenAI APIs.

Add one server. The agent sees every tool, and the server tells it how to use them. Setup for each app is in step 2 above.

Function tools

Your agent calls tools you define, but it has no MCP client.

Load the tool list from /api/v1/tools. When the model calls a tool, send the call to /api/v1/tools/<name> and give the reply back to the model.

REST API

You write the code that decides when to call, and no model is in the loop.

Call the endpoints in the OpenAPI spec from any language.

Questions

When does the voice agent call?
Between 8 a.m. and 9 p.m. in the person's time zone. Set person.timeZone to an IANA zone such as America/Chicago. Without it, calls happen between 11 a.m. and 9 p.m. Eastern. That window is inside 8 a.m. to 9 p.m. in every continental US time zone. Canadian numbers keep to shorter hours: 9:30 a.m. to 8 p.m. on weekdays, 10:30 a.m. to 5 p.m. on Saturday, and noon to 5 p.m. on Sunday. A task for a Canadian number must set person.timeZone, such as America/Toronto. To call at any hour, such as for an urgent shift, set callAnytime: true on the task.
How many times does it try to reach someone?
Up to maxAttempts calls: 3 by default and 5 at most. Each new call waits at least 4 hours and starts inside calling hours. The voice agent leaves one voicemail per task at most. If the person calls the number back, the voice agent picks up the open task on that call.
What does a call cost?
You pay a per-minute rate for connected time only. Ringing and unanswered calls are free. A voicemail is billed from the moment the message starts. A call starts only when your balance covers 5 minutes. A call ends after 30 minutes, or sooner when your balance runs low. The voice agent says it must wrap up one minute before the limit. Each phone number also has a monthly fee.
Can someone listen to a call or take it over?
Yes. Your agent calls create_live_link and posts the link in your chat. A listen link plays the call in the browser with a live transcript. A takeover link also lets a person from your organization take the call, hand it back to the voice agent, or end it. call_operator dials a person's phone into the call instead. The voice agent tells the other person who is joining before the handoff. Your agent can also follow the transcript with watch_call. Every call opens with "This call is recorded and may be monitored."
Can I share a call recording?
Yes. share_recording creates a link to the recording and transcript. Links are redacted by default: a tone replaces personal, payment, and health details. Redaction costs 2.5 cents per call minute, charged once per call. Links last 7 days by default and 30 days at most, and you can revoke one at any time. Use a redacted link for anyone outside your organization.
How do I check that a webhook came from Hollerback?
Each request has a Hollerback-Signature header in the form t=<unix seconds>,v1=<hex>. The v1 value is an HMAC-SHA256 of <t>.<raw body> with your signing secret. The secret is on the settings page. Compute the HMAC yourself and compare. Reject requests older than 5 minutes.
What if my webhook endpoint is down?
A failed delivery is tried 5 more times, after 1, 5, 30, 120, and 720 minutes. That covers about 15 hours. Your agent can also read a task at any time with get_task or GET /api/v1/tasks/<id>.
What does Hollerback run on?

Each service handles one part of a call or of your account.

Twilio
Phone numbers and the carrier network.
LiveKit
Real-time audio between the phone call and the voice agent.
Deepgram
Speech to text.
Cartesia
Text to speech. This is the voice the person hears.
Anthropic
Claude Haiku runs the conversation. Claude Sonnet writes the result.
Neon
The Postgres database.
Vercel
Hosts this site and the API.
Fly.io
Runs the voice agent and the call scheduler.
Stripe
Payments for your prepaid balance.
GitHub
Sign-in.

Technical reference

These files hold the exact details. Agents can read them too: point an agent at the agent guide and it can set itself up.

  • Agent guide/agents.md

    Everything an agent needs in one Markdown file: setup, the workflow, task fields, states, and errors.

  • llms.txt/llms.txt

    The same guide at the path that AI tools check by default.

  • Skill file/skill.md

    The guide as a SKILL.md file, for agents that load skills from a folder.

  • Tool definitions/api/v1/tools?format=anthropic

    Every tool as JSON Schema. Add format=openai, openai-chat, or mcp for other shapes.

  • OpenAPI spec/api/v1/openapi.json

    Every REST endpoint, request, and response.