Archie for an IT review.
For whoever signs off on software at your work. The full document to attach to a ticket is at the bottom.
Twelve answers for a review.
| The question | The answer |
|---|---|
| What is it? | A desktop app on one computer. The agent is a process on that computer, so closing Archie stops it. No Archie server runs it for you. |
| Does it accept connections? | Nothing on the network can reach it. Three things listen on 127.0.0.1 only, and only while in use: a sign-in waiting for the browser to hand you back, the Google Drive picker, and the browser Archie drives during a Computer control job. |
| Does it need admin rights? | Not to install or run it: the Windows installer writes to the user’s own profile. One optional feature, waking the computer for a routine, asks for more. On a Mac it installs a helper that runs as root, after an administrator’s password; on Windows it is a scheduled task under the user’s own account. No driver or kernel extension; nothing runs at startup unless you turn it on. |
| Where does the work go? | Straight from your computer to your AI company, on your account, with your key, and we keep no copy. Two exceptions run through our server: the free starter credits and the plan with the AI included. On either, your agent reads no email or texts; both work on an AI account of your own. |
| Can a hostile email or web page hijack it? | Partly, and this is the one to weigh. The approval gate stops an injected instruction from changing anything in a connected account that is not easily reversible. It does not fully stop a leak through a web address or a search. Since Archie 0.3.5, an address the agent wrote itself waits for a person’s yes, and a searching skill gets only its task. Some routes remain. The long version → |
| Where do its messages go? | Archie’s own window; a chat app, only if one is connected; and your phone, if phone access is on, sealed in a mailbox on our servers. A chat app holds what it carries: Telegram’s cloud chats are not end-to-end encrypted, and Slack and Discord keep content under their own terms. No chat app, or Matrix on a homeserver you run, also works. |
| What do your servers hold? | Your email address and whether you have a current plan. Then, where they apply, the version heartbeat and, if you leave them on, crash and error reports. That is a summary: the whole list is on the Trust page. |
| Is there analytics? | No. Nothing records what you do in Archie or counts what you use. Once every six hours it reports its version and whether it runs on Mac or Windows. After an unexpected quit, the next launch sends the tail of the crash, and repeated failures send their error line. All of it switches off in Settings, under Privacy. |
| Can it run commands? | No. No tool runs a shell command, evaluates code or loads a plugin, and model output never becomes a command line. The built-in tools are a fixed list in the source. A person can add a remote server’s tools by connecting its address; anything it does not mark read-only waits for their approval. |
| Can it use other apps on the computer? | On a Mac, only once Computer control is on, and only the apps the person allows. It never uses a terminal, password keepers, System Settings, Mail, Messages, Calendar or web browsers, whatever the person says. It never types a password or a card you pay with, and anything it could not undo waits for a person. On Windows, Computer control uses websites only. |
| Can it reach inside our network? | Only where the person using it points it. The web tools refuse private, loopback and carrier-grade NAT addresses, checked inside the HTTP client’s own name resolver. Three things can reach local addresses: lights added on the local wifi, an AI model on this computer, and the Computer control browser. That browser is off until turned on, and can then open a local address. |
| Is it signed? | macOS: yes, with an Apple Developer ID certificate, notarized by Apple and stapled, so Gatekeeper opens it without a warning. Windows: no. The installer is unsigned and SmartScreen warns on first launch. Why, and what to check instead → |
| What about our mailboxes? | Google and Microsoft mail connect through OAuth in your own browser; someone who wants only the calendar never grants mail. iCloud, Fastmail, Yahoo, AOL, Zoho and GMX use an app password. A drafted reply goes out when a person sends it or sets a time for it, and the agent cannot set that time itself. |
The document, for the ticket.
Written for whoever has to approve Archie. What it covers:
- A threat model: what Archie does and does not defend against.
- The hosts Archie connects to and the file paths it creates.
- Every OAuth permission it can request, and the operating system prompts it triggers.
- What each AI provider does with what you send, and an API key compared with an enterprise agreement.
- Twenty open points, and the review artifacts we do and do not have, including ones we have not built.
Since 0.2.3, a Mac also asks for an administrator’s password once, to let routines wake the computer, and for Accessibility, in System Settings, before your agent can use the apps you allow with Computer control.
Revision 5, September 3, 2026, covering Archie 0.2.3 · 33 pages. Where it and this page differ, this page and the Trust page are current.
Check us with a packet sniffer.
About ten minutes: install a network monitor, use Archie properly, and compare what you see with the hostnames we publish. Then, on an AI account of your own, block us at the firewall and watch the agent keep working.
When one row stops your review.
Tell us which, or anything Archie does that this page does not describe, publicly if you like. We will answer with what the code does, and say so plainly if Archie cannot satisfy your policy.