We keep what it takes to run your account. We don't keep your voice, your screen or your files. Crash reports and feedback reach us only when you choose to send them.
Last updated September 30, 2026 (hosted voice accounting; iPhone Mirroring and anonymous iPhone feedback retained) · Co-Agent is a Source4AI product of Automated AI Solutions LLC
"Never keep" is not the same as "never touches," and we won't blur the two. If we're paying for your AI (you're a subscriber, or on a free trial), some requests pass through our server on their way to the AI company that answers them, because that is how we pay for them on your behalf. The next section says exactly which ones. We forward them and keep nothing: no copies, no content in our logs.
We do not sell or rent your data, or share it for advertising. The service providers named on this page process only what is needed to provide the features you choose. Our business model is embarrassingly old-fashioned: you pay us for software, and that's the whole revenue plan.
When you choose “Allow hosted voice” in Settings → Privacy or during setup, audio from your spoken question, shared screen images and conversation context travel over an encrypted connection to Co-Agent's hosted voice server and then over a separate encrypted connection to Google's Gemini API. Both services process the content to provide the reply. Co-Agent keeps it only in bounded memory buffers during the session, without saving recordings or conversation content. Turn off hosted voice in Settings to stop this processing. Local dictation remains available without it.
One exception, clearly bounded: during a step-by-step walkthrough you asked for (a guided tutorial where the app points at what to click), the app also takes a screenshot after each of your clicks so it can see where you are and place the next step accurately. Those screenshots follow the same path as the rest of that conversation, and it only happens while a walkthrough you started is active. Pressing the talk keys (or Escape) cancels the walkthrough instantly, which also stops those screenshots.
The server keeps provider credentials private and measures active conversation time, including sent speech, reply audio forwarded and response time; overlapping time counts once. Idle time and failed connections are free. A one-use connection permission expires after one minute; a connection lasts at most 30 minutes and is closed sooner if idle or its reserved minutes run out. Reconnecting uses the same account balance. We retain usage totals and payment references, not the audio, images, transcripts or tool-result content carried by this connection.
We use Google's paid API, which per Google's terms is not used to train their models.
Release the keys and the microphone lane is closed. Silence sends nothing. There is one wake word, and only in one place: if you start a meeting on the phone, saying “Co-Agent” turns that sentence into a question. Nothing else in Co-Agent listens for a name.
The hosted voice server decrypts conversation content in memory so it can forward replies and enforce account limits. It is not a blind relay. Google's connection comes from our server; Google does not receive your computer's network address on this hosted voice connection. Co-Agent sees connection metadata and processes the content without retaining it.
The earlier Blind Private Relay carries an encrypted connection to Google without decrypting it. It is a separate transport and is not used to enforce purchased hosted voice minutes. Live Avatar continues to connect directly to Google, with its separate allowance described below.
Live Avatar lets Co-Agent answer as a realistic talking video face instead of the 3D character. It is an optional beta, available only on accounts we've enabled it for, and only when you pick a live avatar.
Remote Guide is end-to-end encrypted. When one person guides another over a remote connection, their screens and clicks travel encrypted between the two Macs and are decrypted only on the other side. Our relay forwards ciphertext it cannot read.
Classroom is an optional way to share control with another person on your local network. Everything stays between the two Macs, encrypted locally. Nothing goes through our servers.
The Co-Agent phone app is a window onto the Co-Agent running on your own Mac. Everything it shows you comes from there.
Pairing. Your phone and your Mac agree on a key directly with each other. During pairing both screens show the same six digits; comparing them is what proves nobody has slipped into the middle. If the digits differ, don't continue.
Everything between them is sealed. Each message is encrypted on the device that sends it and only opened by the device it's addressed to, with separate keys for each direction. Our relay carries those sealed messages between your two devices because your phone and your Mac usually can't reach each other directly. The relay cannot read them. It stores nothing. Its logs record that a link exists, when it's busy, and how many bytes went by, not a word of what they say, and your IP address is replaced with a scrambled code that changes every day.
This is the one part of Co-Agent where end-to-end encryption is the honest description, because both ends are your own devices.
Your phone does not send us your voice. It turns what you say into text on the phone itself, using Apple's speech recognition, and only the finished sentence travels. Where your iPhone supports on-device recognition we require it; on hardware that doesn't, Apple performs the recognition on its own systems under Apple's privacy terms, not ours.
The phone asks first. Before your first spoken or typed question is sent for an AI answer, the app explains that your words travel through your Mac to Google's Gemini API and asks permission. Talking and chat stay off until you allow it; Settings → Privacy on the phone lets you take that permission back. Demo boards answer on the phone and send nothing. Hosted voice on your Mac also requires the separate permission described above.
What crosses the sealed link: the sentence you said, the answer and its audio, the tiles on your board, files you deliberately choose to share, avatar files you download from your own Mac, and, only if you start a meeting recording, that meeting’s transcript. A meeting transcribes the room, not just you. The transcribing happens on your phone and stays there, but each time you say “Co-Agent” during a meeting the last few pages of that running transcript travel with your question so the answer makes sense in context, and that can include what other people in the room said. Tell them the meeting is being transcribed; some places require it. Files you share land in a folder on your Mac. None of it is stored by us.
Controlling your phone through iPhone Mirroring. When your iPhone is near your Mac and signed in to the same Apple Account, Co-Agent can control it through the Mac's own iPhone Mirroring window instead of our relay. To do that, Co-Agent reads what is on your phone's screen inside that window, on your Mac, and clicks and types there, under the same rules as controlling your Mac: anything that would send, delete, buy, call or change an account waits for your yes, and it never types passwords. Nothing from the Mirroring window goes to our servers; what happens with a picture of it is the same as the next paragraph.
When you ask about your phone's screen. If you ask Co-Agent what's on your phone, or ask it to do something there, your Mac takes a picture of your phone’s screen and sends it away with your question: to Google, or to OpenRouter if an OpenRouter key is the only one you have given it. When we're paying for your AI that request passes through our server on the way, as described above; with your own key it goes direct. It only happens when you ask, and it isn't stored. When Co-Agent acts on your phone, the taps themselves travel over the cable or your Wi-Fi between your phone and your Mac, or through the same sealed relay when you're away from home. (Deciding which thing to tap can use Jev, described above.) Each action waits for your approval, unless you have chosen to allow them for a set time, which you can end whenever you like.
Real work (files, research, business context) is handled by your own agent (Hermes, Claude Code, Codex…) running under your accounts on your machine. None of it routes through us. Actions on your computer require your explicit on-screen approval and are receipted locally, for you.
The Agent Computer that comes with Pro is a private virtual computer on our servers at Amazon Web Services in the US. It is disposable by design: created when you need it, and deleted after 15 minutes with nobody watching or after an hour at most. Nothing on it is kept after that, so don't store anything precious there.
Its screen, your clicks and typing, and anything you paste into it travel between your app and that computer through our server, encrypted (TLS) on the way. We don't watch or record them. We keep only which license it belongs to and when it was created and last used.
The voice demo streams your held-button audio through our server to Google's Gemini API and is not stored by us. Demo computers are deleted automatically minutes after creation. We keep coarse rate-limiting counters by IP for a few hours to prevent abuse.
Only when you choose to send one. Co-Agent never sends a crash report or feedback on its own. If the app quit unexpectedly, it asks you first; nothing leaves your device until you click Send, and Don't send sends nothing at all, not even a note that you declined. You can also tell it to stop asking.
On iPhone, feedback is anonymous and there is no crash reporter. The iPhone app sends only the message you write, its app version and, if you tick the box, your phone model and iOS version. No email, no screenshot, no account and no device identifier go with it: each report carries a fresh random number that is never stored on your phone, so two reports can't be tied to each other or to you. Want a reply? Email us from your own mail app. The iPhone app has no crash reporter of its own; if you share crash data with app developers in iOS Settings, Apple sends us its standard anonymous crash logs. The rest of this section is about the Mac app.
You see it before it goes. You can read the report preview before sending and add your own note. Our server also records when it arrived and the authentication details described below.
A screenshot you attach to feedback in the Mac app is the one time we keep a picture of your screen, because you chose to send it.
Card processing is handled by Stripe; card numbers never touch our systems.
There's no delete button on the website yet. In the iPhone and iPad app, Settings → Privacy → Delete my Co-Agent account opens an email deletion request. Or email [email protected] from the address on your account and ask. A human will delete your account and sign you out everywhere, along with your sign-in history, the names of your computers, your usage counts and Live Avatar minutes, and any crash reports or feedback you sent. To cancel a subscription, do that first from your account page, so you aren't charged again.
Trial-abuse hashes and security audit records remain after deletion to prevent repeated free trials and protect the service. Trial identity records use one-way hashes rather than storing the deleted email in plain text.
Stripe keeps its own payment records under its own terms, and we keep the payment records the law requires for tax and accounting.
Email [email protected] (or [email protected]) to unsubscribe or ask anything about this policy. A human answers.