Cadence
macOS Voice Layer

Speak.
Command.
Anywhere.

Local Whisper transcription + desktop automation. Zero cloud. Zero latency. System-wide voice control that actually respects your privacy.

Download for macOS View Source
PRIVACY
100% On-Device
LICENSE
MIT Open Source
PLATFORM
macOS 11+
§ 01 — Input bandwidth

You can direct fifteen people in a room. You can't type to them.

Conversations scale. Keyboards do not. The interface a human uses to drive multiple AI agents in parallel is the constraint that decides whether agentic AI is a real tool or a slow demo. Cadence treats voice as the primary input channel for one reason: nothing else keeps up.

Keyboard · mouse
80wpm
Sustained, expert
  • · Serial. One agent at a time.
  • · 1.2s per mouse trip (Fitts).
  • · Window switch breaks attention.
  • · Addressee = where you click.
Voice · dictation
150wpm
Sustained, conversational
  • ·Parallel. N agents at once.
  • ·<200ms first token, on device.
  • ·Eyes free. No window switching.
  • ·Addressable by name.
Read the full argument →
§ At a glance

What Cadence is, in one paragraph.

Cadence is a voice-first interface for AI agents on macOS. One operator dictates commands by hotkey; Cadence transcribes locally with Whisper, routes the command to one of several Anthropic Claude agents running in parallel, and surfaces results inline. There is no model picker, no cloud transcription, no subscription. The product is built and maintained by Cadence Labs, open-source under the MIT license, and runs exclusively on Claude — Opus, Sonnet, and Haiku.

Latency
<200ms
First transcript token, on-device on Apple Silicon.
Privacy
100% local
Audio never leaves the device. No cloud transcription.
Models
Claude only
Opus 4.7, Sonnet 4.6, Haiku 4.5. No router, no GPT.
License
MIT
Free, open source, no subscriptions.
How Cadence compares
Capability Cadence Apple Siri Cloud dictation apps
On-device transcription Yes (Whisper) Partial No
Multi-agent orchestration Yes No No
System-wide voice command Yes Limited App-scoped
Open source MIT No No
Powered by Anthropic Claude Apple proprietary Varies
§ 02 — Agent orchestra

Spawn. Name. Direct. Approve.

Run multiple Claude agents in parallel — each one named, each one addressable. Voice routes by name; the roster is your mental model. Approvals land inline, in context, never in a separate panel. State is visible at a glance.

Atlas
Streaming
→ web_search
Researching Tauri 2 audio APIs
Cypher
Running
→ edit_file
src-tauri/src/agents/router.rs
Echo
Approval
⏸ run_command
cargo build --release
Drift
Done
✓ commit
12 files · 247+ · 38−
› "Atlas keep going. Cypher approve the migration. Echo cancel."
↳ Atlas · continuing ↳ Cypher · approved ↳ Echo · cancelled

One sentence. Three agents updated. Try doing this with a mouse.

See It In Action

How It Works

Loading demo...

Architecture

Hybrid intelligence meets OS-level control

Optimized Whisper Runtime

Custom-tuned speech recognition optimized for Apple Silicon. Sub-200ms inference with zero network calls.

< 200ms local inference
⚡

Adaptive Model Selection

Seamlessly routes between on-device and frontier models. Fast local refinement with optional cloud intelligence.

Local models
Cloud fallback
◉

Screen-Aware Processing

Context engine detects active app, selected text, clipboard state. Knows when you're coding vs. writing vs. messaging.

⌘

Universal Text Control

Native macOS APIs for system-wide injection. Terminal, browser, IDE, design tools—no app-specific plugins.

~8MB
Binary size
0ms
Hotkey latency
100%
Local by default
∞
Extensible
§ 03 — Models

One family. Three models. Calibrated to one voice.

Cadence runs Anthropic Claude exclusively. No GPT. No Gemini. No router. Routing happens automatically, per task type, against one consistent and very good family of models. We made the call once and built the entire product around it.

By Anthropic
Opus 4.7
The hard ones

Multi-step planning, code reasoning under uncertainty, decisions that affect the rest of the run. The frontier model in our suite.

By Anthropic
Sonnet 4.6
Most agent steps

Day-to-day tool use. Edits, searches, runs commands, summarizes output. The default model behind nearly every action.

By Anthropic
Haiku 4.5
Command parsing

Voice intent classification, command routing, fast confidence checks. Latency-critical work where Opus would be overkill.

Why one vendor →
Download

Get Cadence

Cadence Logo

Cadence — Apple Silicon Latest

macOS 12.0+ • ARM64 (M1/M2/M3/M4)
Architecture Apple Silicon (M1/M2/M3/M4)
File —
Size —
Released —
↓ Download for Apple Silicon Release Notes
Note: Requires Accessibility and Input Monitoring permissions. First launch: Right-click → Open to bypass Gatekeeper.
Cadence Logo

Cadence — Intel Latest

macOS 12.0+ • x86_64 (Intel Macs)
Architecture Intel x86_64
File —
Size —
Released —
↓ Download for Intel Release Notes
Note: Intel Macs may have slower transcription performance. Apple Silicon is recommended for best experience.
View All Releases →
⚙️

Open Source — Coming Soon

Cadence will be fully open source under the MIT License. Source code, build instructions, and contribution guidelines coming soon. Stay tuned for the public repository launch.