Gpt Live

Real-time voice calls with your OpenCode session, on your ChatGPT subscription.
Say what you want. GPT-Live answers instantly, your session does the work, and you hear how it went.

Install · Usage · Configuration · How it works · Contributing

---

Run /voice and OpenCode picks up. You talk; GPT-Live, OpenAI's real-time voice model, listens and replies as you speak. Every request goes to the OpenCode session on your screen, where your usual agent reads, edits and runs things while you keep talking. When the work finishes, you hear the result. You can keep typing in the same session throughout the call.

The aura, rendered by the plugin: at rest, then you (cyan), then GPT-Live (violet).

Talk while it works. Requests queue into your session. Ask "how's it going?", redirect it mid-task, or say "stop that". Approve by voice. When the session asks for permission, GPT-Live reads the request out and answers with your decision. Picks up where you left off. Each session keeps one voice conversation, so the next call remembers the last one. Hang up by voice. Say "end the call", or press a key. Hands-free audio. Echo cancellation and noise suppression run locally, and other apps' audio is turned down during calls. No API key. It uses the ChatGPT sign-in you already have in OpenCode.

Requirements

• OpenCode 2.0.14 or newer.
• A ChatGPT account signed in to OpenCode: run /connect, choose OpenAI, then "ChatGPT Pro/Plus". Voice availability and usage limits depend on your ChatGPT plan.
• A microphone and speakers (or headphones).
• Linux: ALSA (libasound2), which PipeWire and PulseAudio systems already provide. Turning down other apps during calls uses pactl.

Install

Add the plugin to your OpenCode config (~/.config/opencode/opencode.json for every project, or opencode.json in one project):

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["opencode-gpt-live"]
}

Restart OpenCode. The audio helper for your platform comes with the plugin as a prebuilt package. If it is missing, the plugin downloads it from the matching GitHub release and checks it against the release's SHA-256 checksums before running it.

On macOS, the first call asks for microphone access for your terminal app.

Usage

1. Run /voice, or press ctrl+x v. Without a session open, the call starts in a new one.
2. Talk. For example:
   • "What does this project do?"
   • "Add a dark mode toggle to the settings page, then run the tests."
   • "How's it going?" · "Actually, use the existing theme hook instead." · "Stop that."
3. Hang up by saying "end the call", running /voice again, or pressing ctrl+x v.

While a call is on, the strip above the prompt shows who is talking, what OpenCode is doing and the call keys. The transcript panel on the right shows the conversation and every task sent to your session, with the aura on top. It opens with the call and closes when the call ends.

Commands and keys

Command	Key	What it does
/voice	ctrl+x v	Start or end a call
/voice-mute	ctrl+y	Mute or unmute your microphone
/voice-panel	ctrl+s	Show or hide the transcript panel
/voice-stop		End the call (alias /hangup)
/voice-new		Start a call with a fresh voice conversation
/voice-pick		Choose GPT-Live's voice

Every command is also in the ctrl+p palette. Keys can be changed; see Configuration.

Terminal support

The aura picks the best way your terminal can draw it:

Terminal	Aura
Kitty graphics protocol (Ghostty, kitty)	Anti-aliased image, 30 frames/second
herdr	Image through herdr's pane graphics API
Anything else (tmux, iTerm2, Terminal.app, Windows Terminal)	Half-block characters in true color

Voice, commands and the transcript work the same everywhere.

Configuration

Pass options with the object form of the plugin entry. Every option is optional.

{
  "plugins": [
    {
      "package": "opencode-gpt-live",
      "options": {
        "voice": "juniper",
        "duck": true,
        "keybinds": { "toggle": "<leader>v", "mute": ["ctrl+y", "f9"], "panel": "ctrl+s" },
      },
    },
  ],
}

Option	Default	Description
voice	"cove"	GPT-Live's voice: cove, juniper, maple, spruce, ember, vale, breeze, arbor or sol.
keybinds	{ toggle: "<leader>v", mute: "ctrl+y", panel: "ctrl+s" }	Keys for toggle, mute, panel, stop, new and voice. A key, a list of keys, or false.
panel	true	Open the transcript panel when a call starts.
duck	true	Turn other apps' audio down during calls and restore it afterwards.
instructions		Extra instructions appended to GPT-Live's prompt, such as a preferred language or tone.
voiceAgentInstructions		Extra instructions appended to the voice agent's prompt.
prompts		Replace a built-in prompt: { "gptLive": "<path>", "voiceAgent": "<path>" }. See Custom prompts.
voiceModel	"openai/gpt-6.1-sol"	Model for the background agent that thinks for GPT-Live, as provider/model.
voiceVariant	"medium"	Thinking level for that model.
log	true	Keep a local log of each call (see Privacy).
visual	automatic	Force the aura's drawing method: kitty, herdr or blocks.

Custom prompts

A call uses two system prompts: one for GPT-Live, which does the talking, and one for the voice agent, which thinks and acts for it. Both live as folders of Markdown sections in src/server/prompts/. To change one, copy its folder (or write a single Markdown file), edit it, and point the plugin at it:

{
  "plugins": [
    {
      "package": "opencode-gpt-live",
      "options": {
        "prompts": { "gptLive": "~/.config/opencode/voice/gpt-live", "voiceAgent": ".opencode/voice-agent.md" },
      },
    },
  ],
}

Relative paths resolve against the project and ~ expands to your home directory. {{project}} and {{directory}} are filled in, and HTML comments are dropped. Prompts are read at the start of every call, so edits apply to the next call. If a custom prompt can't be read, the call uses the built-in one and says so in the transcript.

How it works

flowchart LR
  you((You)) <-->|microphone, speaker| host["gpt-live-host<br/>Rust: WebRTC, echo cancellation, Opus"]
  host <-->|WebRTC audio| live["GPT-Live<br/>OpenAI real-time voice"]
  tui["Terminal plugin<br/>aura, transcript, keys"] <-->|stdio JSON| host
  tui <-->|RPC| server["Server plugin<br/>ChatGPT sign-in, call setup"]
  server <-->|control WebSocket| live
  server -->|hand-off| voice["Voice session<br/>thinks for GPT-Live"]
  voice -->|queue, steer, stop, permissions| main["Your OpenCode session<br/>does the work"]
  main -.->|progress, results| voice

• gpt-live-host is a small native program that owns the audio: WebRTC to OpenAI, the microphone and speaker, echo cancellation, noise suppression and Opus encoding. The plugin talks to it over newline-delimited JSON on stdio.
• The server plugin holds your ChatGPT credentials, creates the call, joins its control channel and bridges GPT-Live to OpenCode. The native helper never sees your credentials.
• The voice session is a separate OpenCode session per coding session. When GPT-Live needs to act or look something up, it hands off to this session's agent, which can only talk to your main session: queue or steer work, check on it, read what it did, stop it and answer its permission prompts. It cannot read files or run commands itself.
• Your main session does the actual work, exactly as if you had typed the request.

The architecture guide covers each part in detail.

Privacy

• Audio travels directly between your computer and OpenAI over an encrypted WebRTC connection, under your ChatGPT account and OpenAI's terms.
• Your ChatGPT credentials stay in OpenCode; the plugin reads them from OpenCode's ChatGPT sign-in.
• Each call writes a local log of transcripts and events to ~/.local/state/opencode-gpt-live/calls/ ($XDG_STATE_HOME if set, %LOCALAPPDATA% on Windows). Set "log": false to turn it off.
• The voice conversation is an ordinary OpenCode session, stored like any other.

Troubleshooting

Symptom	Fix
"GPT-Live needs your ChatGPT subscription", or a 401	Run /connect, choose OpenAI, then "ChatGPT Pro/Plus".
403 error when the call starts	Your ChatGPT plan may not include voice, or the voice allowance is used up.
GPT-Live can't hear you (macOS)	Allow microphone access for your terminal in System Settings → Privacy & Security → Microphone.
No audio on Linux	Install libasound2 (Debian/Ubuntu) or alsa-lib (Fedora, Arch).
The aura looks blocky	Your terminal has no image support; the half-block version is the fallback.

Still stuck? Open an issue with your OS, terminal and the call log.

Contributing

Contributions are welcome, from bug reports on a platform we haven't tested to new features. Start with CONTRIBUTING.md; the development guide covers setup, running the plugin locally and testing calls without a microphone.

git clone https://github.com/malhashemi/opencode-gpt-live
cd opencode-gpt-live
bun install
bun run check          # formatting, lint, types, tests
bun run build:native   # the audio helper (Rust 1.91+)

Acknowledgments

Built on OpenCode and OpenTUI. The audio helper uses webrtc-rs, cpal, sonora (WebRTC's audio processing in Rust), Opus and rubato. Third-party licenses are listed in NOTICE.

Disclaimer

This is an independent project, not affiliated with or endorsed by OpenAI or the OpenCode team. ChatGPT and GPT-Live are trademarks of OpenAI. The plugin uses the voice service that comes with a ChatGPT subscription, which OpenAI may change at any time.

License

MIT © M. Adel Alhashemi