Speak into Claude Code's editable prompt, using a local microphone and the Aximo speech engine. Review the words and press Enter yourself. Dictation never sends a prompt automatically.
Source preview: the plugin and native helper are implemented; no downloadable runtime release is published yet. One-command end-user setup remains gated on verified platform builds and microphone testing. See installation.
- Run
/av setup en(Parakeet) or/av setup ru(GigaAM). - Approve the runtime/model download once. Allow your OS microphone prompt when recording starts.
- Run
/av, speak, then choose Stop or run/av stop. - The transcript appears at the cursor. Edit it, then press Enter when ready.
Cancel discards dictation. If a dialog blocks insertion, the transcript stays in memory until you choose Insert or Cancel. Recording stops after 60 seconds.
Once a verified native release is available, inside Claude Code:
/plugin install aximo-voice --marketplace agent-axiom/aximo-voice
The native package is designed to require no Rust, Python, Docker, ffmpeg, or Node setup for users. The current source preview still requires a developer-built or CI-built native helper. Claude Code 2.1.293 is the tested Mods host.
- Audio capture and speech recognition run on your computer
- No API key, cloud transcription account, telemetry, or audio history
- Models download from checksum-pinned Hugging Face snapshots, only after consent
- Aximo uses a private temporary WAV during inference, removed on normal completion
- A crash or forced kill can leave private temporary files; see privacy
- Claude Code receives the inserted draft; it is sent to Claude when you submit it
Use /av for dictation. /avoice and the original /aximo-voice are equivalent
aliases, including every subcommand. Claude Code's built-in /voice is unchanged.
The first version loads its local model for each dictation; it is not streaming ASR.