Cadence keeps Claude Code, Codex and the rest moving. Every question and finished task lands in one place. Answer with a keystroke, or just say it.
Claude Code
Codex
Gemini
Cursor
Copilot
OpenCode
Amp
Grok
Pi
OmpCadence's actual interface running in your browser, with sample agents in their terminals. Click or tap into it, then use 1–9, esc, J K and E, or hover the dots.
Keep the agents you already use. Cadence sits at the edge of your screen and tells you the moment one needs you.
Download for Mac
Drag Cadence to Applications and open it. Allow the microphone, accessibility and input monitoring when macOS asks.
Cadence uses the claude CLI you're already signed into, so there's no API key. Codex, OpenCode, Gemini and the rest join through herdr.
How should I handle duplicate Stripe retries?
Press 1 in the Inbox, or double-tap ⌥ and say what you need. Cadence sends it to the right agent and tells you where it went.
Five terminals, three of them waiting on a yes, and you find out ten minutes later when you happen to tab back. Cadence watches every agent for you, brings you only what needs a decision, and sends your answer back to the right place.
A thin toolbar on the edge of your screen shows one dot per agent. Blue is working, amber needs you, green is done. Open the Inbox and answer in a keystroke.
Hover the agent pill to peek at every agent's title. Click any agent to jump straight to it.
Questions, plan approvals and Allow/Deny prompts become numbered options. Press 1 to 9, or start typing a reply.
When an agent finishes, you get a short recap and suggested next steps instead of a wall of output.
J K to move, E to dismiss, V to talk, S to attach a screenshot. Every answer has a two-second take-back.
Double-tap ⌥ and say what you want, or type it. Cadence knows what each agent is working on, which app you're looking at and which repos you have.
Your request goes to the agent already on that work, or Cadence starts a new one in the right folder. It tells you where it went.
An always-on agent that can see every other agent, read their work, message them, start new ones and close finished ones. Ask "what's left before I can ship?"
Every agent gets a short plain-English title, so your list reads like a to-do list.
Put a whole project's agents away as a set. Their windows close, they keep running, and one chip turns amber if anything needs you.
Director. Ask "what's left before I can ship?" and it answers from every agent's actual work.
Sets. Put a project's agents away. They keep running in one chip that turns amber when something needs you.
The same key types for you everywhere: Slack, Mail, your editor, a terminal. Speech recognition runs on your Mac, so your audio never leaves it.
NVIDIA's Parakeet model transcribes while you hold the key, with a second pass to get names right.
Teach it project names, people and jargon, and set auto-corrections for the words it keeps missing.
A small local model cleans up filler words and punctuation, and adjusts tone per app: casual in Slack, formal in email.
Pressed the wrong number? Every answer waits two seconds, and esc takes it back.
1–9 answer, J K move between cards, E clears one, and Space starts a reply.
Press S, drag over part of your screen, and the screenshot goes along with your answer.
Finished work arrives with a one-line recap and replies you can send with one key.
Claude CodeCodexGeminiAmpCursorCopilotOpenCodeGrokPiOmp
Claude Code runs inside Cadence with real allow and deny, cost tracking and worktree diffs. Everything else runs in herdr terminals that Cadence watches the same way.
Pair your iPhone with a QR code and answer your agents from the couch, the train or the coffee line. Traffic between your phone and your Mac is end-to-end encrypted, and your Mac never opens a port. If your Mac falls asleep, carry a Claude Code agent on in the cloud, and the work comes back to your Mac when it wakes.
One email when the iPhone app is ready. Nothing else.
How should I handle duplicate Stripe retries?
Done. Build green, one deprecation left.
I'm Max. I built Cadence because running a few coding agents at once had turned my day into checking terminals: one finishes, one wants a yes, and I find out ten minutes later.
Before Cadence I founded Tycho, a SaaS tool for sneaker resellers, and shipped it to real, paying customers. Cadence is the tool I wanted on my own desk: every agent's question in one place, answered with one key or a sentence out loud.
It's free, it needs no account, and I read every email that comes in at support@cadencevoice.ai.
Speech recognition and Text Polish run locally. Audio is never uploaded, and recordings aren't saved unless you turn that on.
Routing, titles, recaps and the Director use your own Claude Code login. Other agents talk to their own providers, as they already do.
Cloud Polish and the Intent Engine only run if you add your own Anthropic API key.
No Cadence account and no subscription. Details in the privacy policy.
No. It sits on top of the agents you already run, like Claude Code and Codex, and keeps them moving: it shows you what they need, sends your answers back, and routes new requests to the right one.
It never answers for you. Questions, plan approvals and Allow/Deny prompts wait in your Inbox until you answer, and every answer has a two-second take-back. Which tools an agent may use without asking is up to you, the same as in the CLI.
The Director only starts, messages or closes agents when you ask it to.
Not Claude Code: Cadence runs it directly, with real Allow/Deny, cost tracking and worktree diffs. Every other CLI (Codex, OpenCode, Gemini, Amp, Cursor, Copilot, Grok, Pi, Omp) runs in herdr, an open-source terminal workspace, and opens in a Cadence terminal window. Their questions land in the same Inbox.
Nothing. Cadence is free, with no account and no subscription. Your agents keep using the plans you already pay for: your Claude Code login, or whichever provider each CLI is set up with.
Soon. The iPhone app pairs with your Mac over a QR code, and traffic between them is end-to-end encrypted. If your Mac falls asleep, you can continue a Claude Code agent in the cloud and it comes back to your Mac when it wakes. Join the waitlist and you'll get one email when it's ready.
Speech recognition and Text Polish run on your Mac, and audio is never uploaded. Agents, routing, titles and recaps use your own Claude Code login; other agents talk to their own providers, as they already do. Cloud Polish and the Intent Engine only run if you add your own Anthropic API key. The privacy policy has the details.
Dictation and Text Polish do, once their models are downloaded. Agents, Talk routing and the Director need a connection, because they talk to Claude or your other agents' providers.
A Mac with Apple Silicon (M1 or newer) on macOS 13 or later. For agents, the claude CLI installed and signed in, plus herdr if you want agents other than Claude Code. Dictation works without either.
