Frequently asked questions

Cadence is a native macOS control layer for AI coding agents. It sits between you and your agents, so you can run many of them without keeping any of them waiting.

Every question and finished task lands in one Inbox. Answer in a keystroke, or just talk. A thin toolbar on the edge of your screen shows one dot per agent: blue while it works, amber when it needs you, green when it's done. Cadence also does on-device dictation into any app.

Claude Code is hosted natively: real Allow/Deny approvals, cost, and worktree diffs.

Other command-line agents run through herdr, an open-source terminal multiplexer:

• Codex, OpenCode, Pi, Amp, Grok
• Cursor, Copilot, Gemini, Aider

herdr agents open in a Cadence terminal window with tabs (⌘1–⌘9, ⌘N for a new one), and their questions still land in the Inbox.

No. Cadence uses your existing Claude Code login (the claude CLI) for agents, routing, titles, recaps, and the Director.

An Anthropic API key is only needed for two optional features: Cloud Polish (cleans up dictated text with Claude Haiku) and the Intent Engine (interprets voice commands Cadence doesn't recognize). Both are off unless you turn them on.

Yes. No account, no subscription. The agents you run use your own Claude Code plan, or whichever provider each CLI is set up with.

One place for everything your agents need from you: questions, plan approvals, and tool Allow/Deny requests, shown as numbered options. Finished work arrives as a card with a short recap and suggested replies.

• 1–9 answer, J/K move, E discards, Esc closes
• Space or just start typing to reply; V for the mic, S to attach a screenshot of a region
• Every answer has a 2-second take-back

When it's empty, it says "Nothing needs you."

Double-tap ⌥, then say or type what you want. Cadence decides where it goes: to the agent already on that work, to a new agent, or to the Director. If it can't tell, it asks. You get a short receipt like "Sent to {agent}" or "Answering in Director".

The routing runs on Claude Haiku through your own claude login. It looks at your agents' titles and recent output, the app in front, an optional screenshot, and the folders in ~/Developer.

An always-on Claude session that can see and command every agent. Ask it what's going on across your work, and it can start agents, message them, read their output, and close them.

• AI titles and recaps: each agent gets a 2 to 5 word title, and long finished messages get a one-line recap.
• Sets: "Put away as a set" closes a group of agents' windows, keeps them running, and folds them into one chip that turns amber when something needs you.
• Agents dashboard: Chat and Board views, with branch, worktree, and diffs on each card.
• Peek: hover the agent pill on the toolbar for two seconds to see every title.

Coming soon. An iPhone companion that shows your Inbox and agents is in the works. It isn't publicly available yet.

Hold ⌥ and speak; release and the text is typed into whatever app you're in. The key is configurable (⌘ or a combination). The mic button on the toolbar starts continuous dictation instead.

Speech is transcribed on your Mac with NVIDIA Parakeet, plus a Whisper pass for names. It transcribes while you're still holding the key, so text lands quickly after you let go.

In our measurements: 1.88% word error rate on clean speech, and about 330 ms from releasing the key to text (95th percentile). Background noise doesn't produce phantom words.

Yes. Add words to Vocabulary so they're spelled right, and set up auto-corrections for anything it keeps getting wrong. You can also pick which microphone to use.

Local cleanup of filler words, capitalization, and punctuation, using a small on-device model (Qwen 2.5 1.5B). It's on by default, and you can set a tone per app.

Cloud Polish is an optional alternative that uses Claude Haiku with your own Anthropic API key.

Yes. The default model is tuned for English. For other languages, switch to a multilingual speech model in Settings (Parakeet v3 covers 25 languages, and Whisper models are available too).

Yes, as secondary features. Voice commands can open apps, search, manage tabs and windows, drive VS Code, and run custom commands. The wake word (default "computer") is off by default. See the documentation for details.

macOS 13 (Ventura) or later on Apple Silicon (M1 or newer). There is no Intel build.

• The claude CLI (Claude Code), installed and logged in
• herdr, if you want to run agents other than Claude Code

Dictation works without either.

The speech and Text Polish models download on first run: roughly 2 to 3 GB in total, depending on which models you pick.

Dictation and Text Polish do, once the models are downloaded. Agents, Talk routing, and the Director need an internet connection, because they talk to Claude (or your other agents' providers).

Not planned. Cadence is built on macOS APIs (Accessibility, input monitoring, window management) and is macOS-only.

Download the latest .dmg from the releases page, drag Cadence to Applications, and open it. Cadence checks for updates on its own.

Cadence is signed and notarized by Apple from v0.5.0, so it opens like any other app. If macOS warns about it, you have v0.4.0 or older. Download the latest version from the releases page instead.

macOS ties permissions to how an app is signed, and v0.5.0 is signed differently from earlier versions. System Settings can still show Cadence switched on, but those entries belong to the old version and no longer apply. To clear them:

• Quit Cadence
• Run tccutil reset All com.cadence.app in Terminal
• Reopen Cadence and allow Microphone, Accessibility and Input Monitoring again

Later updates keep your permissions.

• Microphone: to hear you
• Accessibility: to type dictated text into other apps
• Input Monitoring: to notice the dictation and Talk hotkeys in any app

If you attach screenshots in Talk or the Inbox, macOS may also ask for Screen Recording. You can revoke any of these in System Settings → Privacy & Security.

No. Audio, dictation, and Text Polish run on-device. Recordings aren't kept unless you turn on the setting that saves them.

The agent side does. Talk routing, titles, recaps, the Director, and Claude Code agents talk to Claude through your own Claude Code login. That can include what you said or typed, your agents' titles and recent output, the name of the app in front, and a screenshot if you attach one. Other CLIs talk to their own providers.

Cloud Polish and the Intent Engine send text to Anthropic only if you add your API key and turn them on.

There's no Cadence account, and the app doesn't report usage to a Cadence server. It does check GitHub for new releases. See the privacy policy for the details.

On the same Wi-Fi, directly. From anywhere else, through a relay: your Mac dials out (no open ports), and everything is end-to-end encrypted with AES-256-GCM, so the relay only sees ciphertext.

With the iPhone app, you can continue a Claude Code agent in the cloud. Your Mac keeps an end-to-end encrypted copy of each agent at the relay; when you tap Continue in the cloud, a private sandbox picks up that agent's repo and conversation, and you keep working with it from your phone. When your Mac wakes, the files, commits and conversation come back to it. See Continue in the cloud.

On your Mac, in ~/Library/Application Support/com.cadence.app: settings, downloaded models, agent history, and titles.

Accessibility and Input Monitoring are powerful permissions, and agents can change files on your machine. Only download Cadence from the official releases page, and review what your agents ask to do before you allow it.

• Check that Input Monitoring and Accessibility are granted in System Settings
• Quit and reopen Cadence after granting permissions
• Already on but still not working after an update? Quit Cadence, run tccutil reset All com.cadence.app in Terminal, then reopen it and allow access again
• Make sure the speech model finished downloading

• Check that Microphone access is granted
• Pick the right input in Cadence's mic picker
• Bluetooth headsets drop to a low-quality mic while recording; the built-in mic is often better
• Add names and jargon to Vocabulary

• Run claude in Terminal and make sure you're logged in
• For Codex, OpenCode, and other non-Claude agents, make sure herdr is installed
• Check your internet connection

Quit Cadence, drag it to Trash, then delete ~/Library/Application Support/com.cadence.app to remove settings and models. Revoke its permissions in System Settings → Privacy & Security.

• Email: support@cadencevoice.ai
• Documentation: cadencevoice.ai/docs