Agent Client Protocol

CallBro uses the Agent Client Protocol (ACP) to connect AI agents to your meeting chat, notes, and titles.

How ACP works

ACP is a standard for communication between an app and an AI agent. The app opens a session and sends your prompt. The agent streams its answer, reports tool activity, and can request permission for actions.

In CallBro, the agent runs as a separate process on your device and communicates over standard input and output. It can still use a cloud model: where the model runs depends on the agent and provider you choose.

ACP connects CallBro to the agent. MCP connects the agent to tools, including CallBro’s tools for reading and updating your vault. See the official ACP documentation for the protocol itself.

Connect an agent

1. Open Settings → Agents.
2. Find your agent and choose Configure or Finish setup. Use Add agent to browse more agents.
3. If an Install button appears, use it. Otherwise, follow the installation instructions shown in the app.
4. Set up the connection: enter a provider API key for Pi, or follow the agent’s sign-in flow for Codex or Claude Code.
5. Once setup is complete, choose Use as default.

The default applies to new chats, enhanced notes, and generated titles. Existing chats keep their selected agent. Enable Available in chat for other agents you want to use in chat.

Pi with an API key

Pi is CallBro’s default agent. Open its configuration, install it if prompted, then choose a Provider, enter its API key, and select a Model. For a local provider, set its Base URL and model instead; a key is only needed if that provider requires one.

For example, choose OpenRouter, enter your OpenRouter key, and select a model offered by your account. Pi uses that connection for meeting chat and note generation.

Codex or Claude Code

Open Codex or Claude Code, install it if prompted, and use the sign-in button shown under Connection. Complete the browser or terminal steps, then return to CallBro. If you signed in outside the app, choose Check again.

These agents use their own account connection. You don’t need to copy a Pi provider key into their configuration. Choose Use as default when the agent is ready.

Try it on a meeting

Open a recorded meeting and choose Add to chat from its actions menu. This attaches the recording to your current chat, or opens a new chat with your default agent. Then try:

Summarize this meeting in five bullets. Separate decisions from open questions.

List the action items with owners and due dates. Write “not specified” where the transcript doesn’t say.

Draft a short follow-up email with the decisions and next steps. Show me the draft here.

For a note edit, be specific about where to save it:

Add an “Action items” section to this meeting’s note. Keep the rest of the note unchanged.

Review the answer against the transcript before sharing it. In Permissions, Ask before actions lets you review sensitive actions. Always trust this agent allows available actions in its chats without further approval, across vaults, until you revoke trust.

Add a custom ACP agent

Use Settings → Agents → Add agent → Connect custom agent for an agent executable or adapter that supports ACP over standard input and output. A normal chat CLI needs an ACP mode or adapter to work here.

For example, to launch the Codex ACP adapter through your own Node.js installation:

Field	Value
Name	Codex custom
Command	npx

Under Advanced → Arguments (one per line), enter:

-y
@agentclientprotocol/codex-acp

This example requires Node.js 20 or later and npx on your system. If CallBro can’t find the command, use the full path to the executable. Put only the executable in Command, with each argument on its own line. Optional environment variables go under Environment, one KEY=VALUE per line.

Choose Add agent, finish its sign-in, then choose Use as default or enable Available in chat. The built-in Codex option is simpler if you don’t need a custom launch command.

What data the agent uses

The agent works in your active vault. It receives your prompts and meeting context, and can read transcripts, notes, and other vault files through tools. It may also read saved memory when answering.

With a cloud model, the content used to answer your request is sent to that model’s provider. Connected external tools may send data to their services too. Local recording and transcription do not make a cloud agent local; use an agent and model configured to run locally if you want that workflow to stay on your device.

If setup fails

• Command not found: check the executable path. For custom agents, use a full path if the app can’t resolve the command name.
• Sign in required: finish the agent’s sign-in, then choose Check again.
• Connection or model error: check the provider key, Base URL, and model in the agent’s configuration. Under Advanced, use Check connection to retry.

For installation and recording basics, see Getting started.