Skip to main content

What stays and what leaves

Your conversations, records, memory, settings and keys are kept on your computer. Archetype sends no usage data. What leaves it is what you set up:

What the agent may do

The agent asks before it edits, runs commands or sends anything, as your permission mode says. A few rules hold whatever the mode:
  • A project’s config can narrow what the agent may do, never widen it, and a project’s named agent cannot replace a built-in one.
  • An engine’s own settings in a folder (.cursor/, opencode.json and others) are asked about once before the engine uses them.
  • Programs that start other programs (sh, env, npm run, docker …) are asked each time, never kept as Always allow.
In Manage → Permissions:

Secrets

  • Keys are in ~/.config/archetype/env, readable by you only; the app shows a key’s name, never its value.
  • A key the agent asks for is typed on a card, kept out of the conversation, and saved in ~/.local/share/archetype/secrets/. The agent can still read that file, so this keeps a key out of the record, not from an agent you do not trust.
  • Codex and the other engines get only the keys they sign in with, and those you list in passEnv.
  • Memory refuses text that looks like a key.

Plugins

Each plugin’s processes run in your system’s sandbox. By default a plugin reaches only its own folder and data, and no network. Its manifest declares what more it needs: the network, or your files. ~/.ssh, ~/.gnupg, cloud credentials and Archetype’s own config and data stay closed to it whatever it declares. Where there is no sandbox (Windows, for one), only Pairful’s plugins run.

A company’s host

A company’s host keeps the company’s conversations, projects and knowledge. Whoever runs the host can read what it keeps. Your own accounts connect in your own Archetype, not on the company’s host, and your own subscription runs only your requests there. See Archetype for Teams.

Report a vulnerability

Email dev@pairful.ai, not a public issue. Say the version (archetype --version), the system, the steps to reproduce it, and what an attacker gains. We answer within 5 working days, and say within 15 whether it is a vulnerability and how serious. We fix it in a release and publish an advisory, crediting you unless you ask us not to. For other problems, see Support.