From the LYNS family · Windows · v1.0.0

The voice of Claude Code.

LYNS Voice is a hands-free desktop orchestrator for Claude Code. Speak with Oracle: your words go in, working code and a spoken reply come out. Every word is processed on your machine.

Download for Windows View on GitHub

100% on-device speech
No cloud, no accounts
MIT licensed
macOS on the way

Live render: the same Three.js orb that ships in the app.

The loop

One conversation loop. Zero keyboard.

LYNS Voice turns Claude Code into a conversation: it hears you, works, answers out loud, and listens again.

01
Speak
Tap to talk, or just say Oracle with the optional wake word. The mic meter confirms it hears you.
02
Auto-send on pause
Voice activity detection notices when you stop talking and sends the request. No hotkey, no submit button.
03
Claude works
The claude CLI runs a real coding session in your project. The orb thinks; the activity panel streams every tool it uses.
04
Spoken streaming reply
Replies are spoken sentence by sentence as they stream, with gapless neural synthesis. The orb flares on every word.
05
Back to listening
The loop rearms itself. Interrupt anytime with Escape, or keep talking hands-free for the whole session.

Barge in anytime: EscapeOptional wake word: just say "Oracle"

Features

A command center you talk to.

On-device speech

Local Whisper transcribes you. Kokoro neural TTS answers in-process, with native Windows SAPI as an option. No audio ever leaves the machine.

Whisper STT · Kokoro TTS

The reactive orb

A holographic Three.js orb with four live states: Idle, Listening, Thinking, Speaking, plus mood tints from Claude itself. Three themes: cyan, aurora, ember.

You are looking at it

Multi-session conductor

Alt+N spawns parallel Claude sessions, each in its own panel. The primary voice session can propose and steer a whole fleet.

Alt+N · fleet cost meter

A real terminal

Per-session terminal views over genuine ConPTY PowerShell, with echo, ANSI, and working resize. Watch Claude work, or take the wheel yourself.

ConPTY · per session

Floating diff viewer

Every Edit and Write that Claude makes streams into a floating, syntax-highlighted diff panel, live as it lands.

Alt+D

Your Library, inherited

The MCP servers and hooks you already configured for Claude Code appear automatically, visible and one-click toggleable per session.

MCP · hooks

Mini mode

Shrink to a 180 by 180 always-on-top orb. Park it in a corner of your screen and keep the conversation going while you work.

PiP · always on top

Guardrails, not shortcuts

Claude runs with an explicit tool allowlist and a denylist, never bypassPermissions. Every tool call is shown in the activity panel.

201 unit tests · 16 e2e

Local first

Your voice stays home.

The only things that ever leave your machine: the claude process doing its job, and a one-time, checksum-verified download of the speech models. That is the whole list.

Local Whisper STT

Speech recognition runs in-process on your CPU with whisper.cpp. Your voice is transcribed on the machine it was spoken to.

In-process neural TTS

Kokoro synthesizes the reply inside the app. No cloud voices, no audio uploads, no per-character bills.

Strict CSP inside the app

The app's own webview allows no remote origins at all. There is no CDN, no telemetry, and no analytics beacon to phone home.

No cloud, no accounts

No signup, no API key stored. LYNS Voice drives the claude CLI you already have, with the login you already did.

LYNS Voice speaking a reply: the holographic orb flares mid-word over the tactical HUD, with the live caption, a session terminal view, and the Library of MCP servers and hooks — Speaking state: Oracle reads the reply aloud while the session panel streams the work and the Library shows every MCP server and hook the session inherited.

Getting started

Installed in one sitting.

Requirements

Windows 10/11 x64 (macOS is on the way)
The claude CLI installed and on PATH
About 600 to 800 MB RAM with the speech stack warm

Install

Download the installer from GitHub releases
Run it, pick a project folder, say hello
Speech models download once, checksum-verified

LYNS Voice_1.0.0_x64-setup.exe

About SmartScreen

The installer is not yet code-signed
If SmartScreen appears: More info, then Run anyway
Checksums are published with every release

Download

Talk to your terminal.

Free and open source under MIT. Windows today, macOS on the way.

Download v1.0.0 for Windows Source on GitHub

FAQ

Questions, answered.

Does my audio go to the cloud?

No. Speech recognition (local Whisper) and speech synthesis (Kokoro, or native Windows SAPI) both run on your machine. The only network traffic is the claude process doing its normal work and a one-time, checksum-verified download of the speech models.

Do I need an API key?

No key is stored by LYNS Voice. You need the claude CLI installed and authenticated; LYNS Voice drives it with the login you already have.

Why does Windows SmartScreen warn me?

The v1 installer is not code-signed, so SmartScreen shows a one-time warning. Choose More info, then Run anyway. Checksums are published with every release so you can verify the download.

Mac or Linux?

Windows today. A macOS port is on the way: the speech stack, terminal layer, and UI are already cross-platform. Linux is not planned for v1.

How is this related to the LYNS IDE?

Same family, same Oracle. LYNS Voice is a standalone product at lyns.app; the LYNS Managerial IDE lives at lyns.dev. You do not need one to use the other.

Can it run multiple Claude sessions?

Yes. Alt+N spawns background sessions in parallel, each with its own panel, model, and effort level, and the fleet cost meter tracks spend across all of them.

What does it cost?

The app is free and MIT licensed. Claude usage is billed through your own Claude Code plan or key, and the app shows live token and cost telemetry per session.