Skip to main content
Glama
Agent voice call MCP

Agent voice call MCP

by daure
README.md
# Voice Call

Local, MCP-driven voice conversations between coding agents and you.

[![Release](https://img.shields.io/github/v/release/daure/voice-call)](https://github.com/daure/voice-call/releases/latest)
[![Build](https://github.com/daure/voice-call/actions/workflows/release.yml/badge.svg)](https://github.com/daure/voice-call/actions/workflows/release.yml)

[Install](#install) · [Take a call](#take-a-call) · [MCP](#connect-an-agent) · [Reference](docs/reference.md)

Agents use Model Context Protocol (MCP) to start calls with a briefing and questions, query call status, and retrieve transcripts and file activity. Answer on your desktop to discuss decisions with an OpenAI Realtime voice assistant. The voice assistant can query project files through read-only tools, and the calling agent receives the conversation when the call ends.

## Features

- Answer, reject, and end calls from a desktop window.
- Discuss project files through read-only search and file tools.
- Review the transcript and file activity after a call.

## Screenshot

![Example voice call with the agent's initial briefing and a conversation transcript](docs/screenshots/voice-call.png)

<sub>Illustrative conversation — discuss the agent's briefing and review the live transcript.</sub>

## Install

**Ubuntu 24.04+ on x86_64**, with a desktop session, microphone, and speakers. Node.js and Electron are bundled.
Answered calls require an OpenAI API key with billing and Realtime access; usage is billed separately from ChatGPT subscriptions.

```sh
installer="$(mktemp)"
curl --proto '=https' --tlsv1.2 -fLsS \
  https://github.com/daure/voice-call/releases/latest/download/voice-call-installer.sh \
  -o "$installer" && sh "$installer"
rm -f "$installer"
export PATH="$HOME/.local/bin:$PATH"
voice-call doctor
```

Keep `~/.local/bin` on your PATH. If the doctor reports "No usable sandbox", run `voice-call setup-sandbox` and repeat the check.
Sandbox setup uses sudo to install a directory-scoped AppArmor exception; anyone who can replace a matching executable inherits it.
See [installation, updates, and sandbox details](docs/reference.md#install).

## Take a call

1. Connect your agent through MCP and ask it to call you with a self-contained briefing.
2. Choose **Answer** to talk or **Reject** to decline.
3. Choose **Hang up** to return the transcript and keep the review window open, or say goodbye to end the call and close the app.

The voice assistant can inspect files but cannot execute commands or change them.

## Connect an agent

Configure your MCP client to run `voice-call mcp` over stdio, with `OPENAI_API_KEY` in its launch environment:

```json
{
  "command": "voice-call",
  "args": ["mcp"]
}
```

The MCP process needs access to your desktop and audio session. Set its tool execution timeout above the requested call duration; calls default to 15 minutes.
See [OpenCode configuration, voice selection, and session-relative file access](docs/reference.md#mcp-configuration).

## Privacy and scope

Briefings, audio, and requested file contents go to OpenAI. Grant access only to project directories you permit the voice assistant to read, and keep secrets outside that root.
Transcripts and file activity live in memory and are lost when MCP restarts. See [file boundaries](docs/reference.md#read-only-voice-tools) and [call limits](docs/reference.md#limits-and-failure-handling).

## Documentation and contributing

- [Development and local installation](docs/reference.md#development) — Node.js 22.20+
- [MCP tools and call results](docs/reference.md#mcp-tools)
- [Verification](docs/reference.md#offline-verification) · [Releasing](docs/reference.md#release)
- [Issues](https://github.com/daure/voice-call/issues)