ThunderPhone v3 is live.Self-serve, from 2¢/min.Read the announcement

Facts verified against published sources on .

Slack

You can run a ThunderPhone agent from Slack. A message starts a call, the call plays out live in a thread, and a reply in the thread steers the agent. When the agent needs a decision, it asks in the thread and waits for your answer. When the call ends, the summary and outcome land in the same thread.

ThunderPhone does not host this bot. You run a small server of your own that connects a Slack app in your workspace to the ThunderPhone public API. The Run calls from Slack guide shows every piece, and an open-source starter bot at github.com/autophonix/thunderphone-slack-bot implements it, ready to fork and deploy.

What the bot does

Start a call from a message. DM the bot or @mention it in a channel with a task, such as "call Comcast and get the outage ETA for account 4471". The bot looks up the number, since ThunderPhone dials only the number it is sent, and places the call with your message as the call's context, so one general agent can handle every task.

Follow the call live. What the agent and the other person say appears in the thread every few seconds.

Steer by replying. A reply in the thread while the call is live reaches the agent privately as a call whisper. The agent works it into its next turn, and the other person never hears it.

Answer the agent's questions. When the agent needs your decision, it posts the question in the thread and tells the other person it is checking. Your next reply is its answer.

Read the result. When the call ends, the bot posts how it ended, the summary and any extracted data.

The same pattern works for Microsoft Teams, Discord or a private web chat page. Only the chat side changes.

How it works

The bot uses these parts of the public API:

  • Starting calls. POST /v1/call places the call, with the Slack message as context and an idempotency key built from the Slack channel and message, so a retried Slack event does not dial twice.
  • Live turns. A webhook endpoint subscribed to telephony.turn delivers each line of the conversation. Delivery is at least once and can arrive out of order, so the bot deduplicates on the event ID and orders lines by position.
  • Steering. POST /v1/calls/{call_id}/whisper sends a thread reply to the agent, from 1 to 1,000 characters.
  • Questions. A custom tool such as ask_owner, created as an API connection, calls the bot. The bot posts the question and holds the request open for a reply, up to 180 seconds. For a longer wait, the tool returns at once and the answer arrives later as a whisper.
  • Results. telephony.complete carries the call's status, end reason, summary and extracted data.

The bot keeps one mapping: which Slack thread belongs to which call. Every webhook and tool request is signed, and the bot verifies ThunderPhone's signatures as well as Slack's.

Setup outline

  1. Create an API key under Organization Settings → Keys, and pick an agent and a number that can place outbound calls.
  2. Turn on call recording for the agent, which live turns need, and Generate a call summary, which the end-of-call summary needs.
  3. Create a Slack app in your own workspace and receive messages through Slack's Events API. The starter repository includes an app manifest with the scopes and events it needs.
  4. Deploy the bot to a host with a public HTTPS address, such as Railway, Fly.io or Render. Fork the starter, or have a coding agent build one from the guide. Zeus, the assistant in the ThunderPhone app, can set up the agent's side and write starter code.
  5. Add a webhook endpoint pointing at the bot and subscribed to telephony.turn and telephony.complete. Turn events only go to endpoints that list them.
  6. Create the question tool as an API connection and link it to the agent. Set its spec.timeout through the API, up to 180 seconds; the dashboard editor has no timeout field yet, and without one the tool gives up after 20 seconds.
  7. Limit who can place calls. The starter only accepts Slack users on an allowlist.
  8. Test a full call, a steering reply, a question with an answer, a question nobody answers, and a call that ends while someone is replying.

Prompt guidance

The prompt should say when the agent asks you, how it phrases the question, and what it tells the other person while it waits. For example: use ask_owner with a short question when you need the owner's decision, tell the person you are checking and that it may take a minute, and offer to take a message or arrange a callback if no answer comes.

Keep the context you send with each call to what the agent needs for that task. Text from the call, such as what the other person said, is untrusted, so the bot should escape it before posting it to Slack.

Other ways to get calls into Slack

If you only want notifications, a webhook endpoint subscribed to completed-call events can feed a small service that posts a summary to a channel. See REST API and webhooks. Separately, dashboard alert rules can post to a Slack channel through an incoming webhook URL when a call metric crosses a threshold.

Operating limits

A Slack message does not replace human handoff. A post creates visibility, but it does not establish that a person saw it or took responsibility within the call. Keep transfers or another staffed escalation route for cases that need immediate action.

While a question tool is waiting, the agent gives the other person short updates and does not respond to what they say until the tool returns. There is no hold music. Give the agent only the tools its calls need, and test what it does when nobody answers. The underlying action pattern is described in function calling. Outbound calls placed from Slack follow the same rules as any other outbound call.