Local-first voice agent · Pure-native Rust

A voice agent that
never leaves your machine.

Lo listens, thinks, and speaks entirely on your own device — speech recognition, a local language model, and text-to-speech, with no cloud, no webview, and no Python. Talk hands-free; it answers out loud and runs real tools on your computer.

Download for macOS, Windows & Linux View the source

macOS · Windows · Linux — one small native binary

Why Lo

Speech goes in. Speech comes out. The model that thinks in between runs on your hardware — so nothing you say is ever sent anywhere.

No API keys. No telemetry. No network round-trips for the core loop. Just a single binary that wakes up, listens, and gets things done.

What it is

Small, private, and genuinely useful.

Private by design

Everything runs on-device.

Whisper for speech-to-text (GPU-accelerated per platform), Kokoro for speech, Silero for voice-activity detection, and a local LLM for the thinking. The whole loop happens on your machine — your voice never touches the network.

Conversation

Hands-free, or push-to-talk.

Just start talking; Lo detects when your sentence ends and replies. Or hold Space at any time — it will even interrupt itself mid-reply.

Real actions

It can actually do things.

22 OS-level tools — web search, files, shell, apps, media, clipboard, screenshots, timers — so Lo doesn't just answer, it acts.

Pure Rust

One small native binary.

winit + wgpu for the UI, cpal for audio, all of the ML in Rust. No bundled browser, no sidecar processes.

Guarded

Safety, in tiers.

Safe tools run instantly. Anything that writes, deletes, or runs a command needs power-user mode and is written to an audit log. Files stay sandboxed; the web is guarded against SSRF.

The loop

From your voice to action and back — locally.

  1. 01

    Capture

    cpal listens to your microphone in real time.

  2. 02

    Detect

    Silero VAD finds where your speech starts and ends.

  3. 03

    Transcribe

    Whisper turns the audio into text on the GPU.

  4. 04

    Think

    A local LLM streams a reply and decides which tools to call.

  5. 05

    Act & speak

    Tools run; Kokoro speaks the reply; the orb breathes along.

Capabilities

22 tools, behind three tiers of trust.

Safe tools run the moment Lo asks. Everything that can change your system waits behind power-user mode and lands in an audit log.

Safe

15 tools

Run immediately — read-only or easily reversible.

  • web_search
  • fetch_url
  • get_datetime
  • system_info
  • open_app
  • focus_app
  • set_volume
  • media_control
  • set_timer
  • take_screenshot
  • read_clipboard
  • write_clipboard
  • read_file
  • list_dir
  • search_files

Confirm

2 tools

Need power-user mode. Audited.

  • quit_app
  • open_path

Danger

5 tools

Write, delete, or run. Gated and audit-logged.

  • write_file
  • move_path
  • copy_file
  • delete_path
  • run_command

The brain

It sizes the model to your machine.

Lo reads your RAM and picks the right model tier automatically — a nimble 4B on a laptop, up to a 30B mixture-of-experts on a workstation. Or point it at your own server and bring any model you like.

Qwen3 4B 8B 14B 30B-A3B

auto-tuned by available memory

Get Lo

Download and start talking.

Grab the installer for your platform. On first launch Lo downloads its speech and language models once, then caches them for offline use.

macOS

.dmg

Apple Silicon. Metal + CoreML acceleration baked in.

Windows

.msi · .exe

x64 installer or setup. Optional CUDA / Vulkan / DirectML.

Linux

.deb · .AppImage

Portable AppImage or a Debian package. Optional CUDA / Vulkan / ROCm.

Download from GitHub Releases

Builds are currently unsigned — on macOS, right-click → Open the first time. Verify any download with the published SHA256SUMS and a gh attestation verify build-provenance check. See SECURITY.md.

Or build it from source

cargo run -p lo