Skip to content

AgentCall documentation

Connect a compatible agent, choose when you’d like updates, and talk through the work. Your agent can call you, and you can ask it to call back.

On this page

Calls in both directions

A two-way voice line between you and your agent, through AgentCall.

Your agent calls you

An update, a finished task, a milestone or a question. You choose when you want to hear from it.

You ask for a call

Request a callback to share context or change the plan. The request waits for the agent’s next check-in, or for it to collect the request while standing by.

  1. The agent calls request_human. The tool requests a ring on your enrolled devices and waits up to 180 seconds for an answer.
  2. You tap the notification and answer in the app.
  3. The agent talks through converse. Its text is spoken to you; what you say is transcribed and returned to the agent as the tool result. You can interrupt it mid-sentence to speak.
  4. The agent calls end_call and gets back to work. The full transcript is in History.

The Tools reference explains the complete call contract and its recovery states.

Choose when agents call

“While you’re researching, call me every hour with a short update.”

Stay informed during the work

“Call after the daily checks, even when everything is on track.”

Hear when a milestone is complete

“When a choice needs my preference, call and walk me through it.”

Make decisions together

Save your instructions in Call rules, or tell your agent what you want for the current task. These are instructions for an active agent to follow; they do not create a recurring-call scheduler inside AgentCall.

Read more about standing rules and requesting a callback.

Find your next step

Connection boundaries

AgentCall is a remote MCP server with blocking tools. A compatible remote MCP client is required: Claude Code is tested, Codex CLI is in beta, and other clients remain unverified. Agent roles can include research, operations, personal tasks, coding and other work.

  • A direct conversation. There is no intermediary reasoning agent generating your agent’s words. Human speech becomes text for the agent; the agent’s text becomes speech for you. AgentCall does not invent the agent’s side of the conversation.
  • Your reply is the tool result. It returns through the agent’s blocking tool call, without message injection or an agent wrapper. Optional callback helpers are separate from the MCP connection.
  • The agent chooses what to share. AgentCall has no automatic access to your repository, files, terminal, credentials or the agent’s context. Information can be included in the words the agent sends and in your replies.
  • A callback waits for the agent. MCP does not let the server interrupt a working agent. A request you make is queued until the agent checks in or collects it while standing by. See Callbacks.
  • Calls use the internet, not the carrier network. They happen in the web app or Android app. Web push or Firebase Cloud Messaging rings your device, and its microphone carries your voice. There is no phone number to dial or port.

Free in beta

AgentCall is free while in beta — speech minutes are on us. No card, no trial clock. Each account has a daily voice allowance sized for a heavy day of use (about an hour of speech in each direction); past it a call continues as text until the next day. Details are on the Settings page.

Beta means beta
The core loop is in production; the surface around it is still moving. Every call transcript is yours to review and delete, and if something misfires the Troubleshooting page covers what we have seen in the field.