Talk to your coding agent.
Like you'd talk to a person.
Larmor gives Claude Code, Codex or any AI agent a voice. And ears.
curl -fsSL https://larmor.dev/install.sh | bash
Free while in beta · Apple Silicon Macs · runs on your machine
Why voice
Stop typing at it.
Start working with it.
Today you write a prompt, wait, then read a wall of text to find out what happened. With Larmor you say it, carry on with what you were doing, and it talks to you when something matters.
- 3× faster to say.People speak at around 150 words a minute and type at around 40. Context you'd never bother typing comes out naturally.
- Nothing to babysit.It speaks up when it's done or when it's stuck. Every minute it's working is yours, not the terminal's.
- Steer it mid-sentence.Heading the wrong way? Say so. It stops, listens and changes course instead of finishing the wrong thing.
Works with
Bring the agent
you already use.
Larmor runs entirely on your Mac: it hears you, transcribes you and speaks back without sending a second of audio anywhere. One command sets it up. No account, no API keys.
- Claude Code
- Claude Desktop
- Codex
- Gemini CLI
- Any AI agent
curl -fsSL https://larmor.dev/install.sh | bash
Open source, MIT. Read the code on GitHub before you run it.
- 01
Run the line
It sets everything up for the agents you pick, then downloads the voice models (about 3.7 GB) with a progress bar.
- 02
Open a new session
In Claude Code, type
/larmor. In Codex,$larmor. In any other agent, ask it for voice mode. - 03
Say hello
Your agent answers in a voice. From here on it's a conversation.
Questions
Good questions.
Is my audio private?
Nothing leaves your Mac. Speech recognition (Parakeet) and the voice (Chatterbox) run on your machine, and your mic is only open while voice mode is on. There's no account and no server of ours in the loop. Transcripts are kept in a local log on your Mac, nowhere else. The only things we ever receive are your email when you install, and anything you write in "Send feedback". The code is open source, so you can check all of this yourself.
Is it open source?
Yes, under the MIT licence. The code is on GitHub: the installer, the speech engine, and the one small endpoint that receives your email. The README lists every file the installer touches on your Mac.
Which agents work with it?
Any AI agent that can add MCP tools. The installer sets up Claude Code, Claude Desktop, Codex, Gemini CLI and Antigravity for you, and every agent hears what you say while it works. Claude Code and Codex also get hooks, so they can't wander off mid-conversation and your words reach them within seconds.
How fast is it?
Fast enough to feel like a conversation. You usually hear the first word back in well under a second; the rest depends on how quickly your agent thinks.
What happens when I interrupt?
It goes quiet the moment you start talking. Keep going and it stops and listens; nothing you said is lost. A quick "mm-hmm" and it picks the sentence back up.
What does it cost?
Nothing while it's in beta. Everything runs on your own Mac, so there's no usage to meter.
What do I need?
An Apple Silicon Mac (M1 or later) on macOS 14 or newer, about 6 GB of free disk, and a coding agent. Larmor uses about 3 GB of memory while it runs, so 16 GB is comfortable and 8 GB is tight. Windows and Linux aren't supported yet.
How do I uninstall it?
Run ~/.larmor/app/install.sh --uninstall to stop it and remove it from your login items, then claude mcp remove larmor -s user. The README has the steps for other agents and for freeing the disk space.