Co-Agent is a voice-first AI that lives on your computer — Mac or Windows — and brings its own. Hold a key, say the thing, and watch it happen on your screen or on its private cloud desktop.
Signed & notarized · macOS 13+ · Windows finishing beta
These are not videos. The voice below is the real engine that ships in the app, and the computer is a real cloud desktop minted for you when you press the button.
Press start and a REAL disposable cloud desktop is created for you — then hold the talk button and boss it around. It points, it opens sites, you watch it happen on the machine below.
A real Linux desktop, born when you click, wiped three minutes later. Nothing here is a video.
Yours — where it points, guides, and drafts. And its own — where it does the messy work without touching your stuff.
Hold a chord, ask anything about what's on your screen. It answers out loud in seconds and points at the exact pixel it means — “it's right here.”
A private cloud desktop that boots on demand and sleeps when idle. Your agent browses, downloads, installs, and works there — you watch live in a floating window.
“Show me how” walks you click-by-click with on-screen arrows. “Take me there” clicks for you, step by step, stopping the instant you touch the mouse.
Do the task once while it records. Every click becomes a step; out comes an SOP document and an agent-readable file — your process, captured and repeatable.
Heavy work is handed to YOUR agent — Hermes, Claude Code, Codex, whatever you run. Co-Agent is the voice and hands; your agent is the brain that knows your business.
Draw on your screen while you talk — the marked-up screenshot goes straight to your agent with your words. “Fix this part” finally means something to a machine.
Voice reaches Google through our relay, so Google never sees your IP or location — and the relay can't read the encrypted stream. No single company holds both who-you-are and what-you-said. On by default.
Six moments from a normal day with Co-Agent. They play themselves — or click around.
An assistant on your computer should be deaf and blind until you say otherwise. So it is.
What we collect: your email address. That's the list. No trackers, no analytics, no ad pixels — on this site or in the app — and we never sell or share data, period. Emails are used for your account and occasional offers from us, with a working unsubscribe. Read the whole thing in plain English: the one-page privacy policy.
This demo page: holding the talk button streams your voice to our server and on to Google's Gemini Live API to generate the reply; release the button and the stream stops. Drawing on the page can also turn the microphone on (with your browser's permission and a clear on-screen indicator) so you can talk while you circle things — press Esc and it's off. We don't store demo audio. Demo computers are deleted automatically after three minutes.
Voice included on every subscription. No metering anxiety — typical daily use costs us pennies, so we don't nickel-and-dime you.
Mac and Windows. One signed download, no terminal, no setup scripts. macOS 13+ today; the Windows build is finishing beta — join the list and you'll get it first.
Google's Gemini Live models power speech understanding and the voice itself. Subscriptions include voice through our pooled keys with per-session single-use tokens; lifetime licenses run on your own free Gemini key. Heavy tasks run on your own agent — Hermes, Claude Code, or Codex.
Only with your approval. Any real input on your machine raises an Allow/Deny card stating exactly what it wants to do, and every executed action is recorded in a local receipt you can audit.
A private Linux cloud desktop that exists when you need it. Your agent can browse, download files, install apps, and run long jobs there instead of on your machine. It boots in seconds from a warm pool, sleeps when you're not using it, and its files are disposable by design.
We don't train anything on your data, and voice turns are processed through Google's paid Gemini API. We keep operational metering (token counts), not your content.