Abby narrates Claude Code and Codex answers through a speech model on your own machine —
totally free: no cloud, no API key, no account, no audio ever leaving the computer. Unplug
the network and she carries on talking.
She reads a long answer as its ## TL;DR rather than in full, and Claude is told to write one.
How she decides what to say →
Easy install
Windows only. Paste this into Claude Code:
Clone https://github.com/TraxData313/claude-voice here, then follow docs/tour.md
in it: set it up, turn it on, and give me the spoken tour.
Or in PowerShell, no clone:
powershell -c "irm https://raw.githubusercontent.com/TraxData313/claude-voice/main/install.ps1 | iex"
Either way it becomes a program: Abby for Claude in the Start menu, an Uninstall in
Settings → Apps, no admin rights, fine with Smart App Control. Also as a Claude Code plugin.
What it does →
What it needs
| |
|---|
| Windows | there is no macOS or Linux build. Python comes with it: a private, signed python.org copy |
| Nothing else, for Pocket | the default engine: the CPU, a few hundred MB, Abby in English |
| An NVIDIA card, for Qwen | ~4 GB of video memory (an estimate) and 3 GB of disk; reads every alphabet. Not with Smart App Control on |
Three engines, and what each costs →
Usage cheatsheet
/voice <thing> in Claude Code. Mind the space: /voice on, never /voice_on.
| |
|---|
on / off | the switch. off leaves the engine warm, so on is instant |
set abby | change voice. Any unambiguous substring: set ab works |
engine pocket | which engine speaks: qwen, pocket or breeze. Engines → |
repeat | say the last answer again. repeat-all for the whole thing |
stop / pause / play | cut it off and empty the queue · hold her mid-word · carry on |
volume 40 | how loud, 0 to 100 — audible mid-sentence |
narrate transcript on | also hear the short lines between tool calls; then hooks install. Off by default |
panel | the window: who speaks, pause, skip, volume, history. All of it → |
update | is there a newer one? It only ever looks when you ask |
help | every command, grouped. All of them → |
The voices
Three ship with it, all the author's own — original characters, borrowed from nobody:
| | |
|---|
| Abby | Cute, slightly nerdy, young and warm. Calm, never in a hurry — for careful work, tricky debugging, and being talked through something gently. |
| Neya | Soft-spoken, emotionally intuitive, young and calming. Grounded and gently reassuring — for thoughtful deep dives, empathetic listening, and navigating complex ideas at a comfortable pace. |
| Max | Brave and driving, a trainer's energy. Punchy lines, counts off what is done, pushes for one more — for grinding through a long list. |
▶ Hear all three — GitHub will not play
audio inside a README. Cloning your own, and whose voice you may publish →
With thanks to Genndy Tartakovsky, whose Samurai Jack is why they look the way they
do. The characters are ours; the debt is to the style, and to the spirit of the thing.
Where the detail lives
Freely given
- Public domain — no license, no strings, no permission to ask (The Unlicense).
Use it, share it, change it, sell it. "Freely you have received; freely give." (The bundled
Harmony library keeps its own MIT notice in
lib\.)
- Want to help? Give feedback and report bugs.
- No donations — this is a hobby, done for fun and out of good will; I want to keep money
out of it. "For the love of money is the root of all evil."
- If you still insist on thanking me somehow — visit my GitHub acc and read my top pinned
