Security
Security model
The boundaries that keep agents, credentials, computers, and paired devices scoped.
Seira is designed around a local owner, explicit capabilities, and narrow process boundaries. Local-first does not mean risk-free: an agent with tool or computer access can still act on the resources you grant it.
Core boundaries
- The harness owns agent processes and listens on loopback.
- The renderer receives configured-or-not credential state, not secret values.
- Provider-specific credentials are injected only into the process that needs them.
- Bots start at Full access, as in T3 Code: each engine runs in its own most permissive mode. Choose Ask for approval on a bot or thread to turn risky actions into approval cards. Real questions and missing credentials always come to you.
- Local computer control is a separate opt-in and per-bot selection.
- Paired phones use a default-deny API allowlist and per-device bearer tokens.
- Cloud and VPS backends are validated before reuse.
- The app downloads the project's public announcements file from GitHub, at most hourly and with nothing of yours in it, and may show one of its cards (product news, a tip, a sponsor's offer) once per person. The cards are plain text with fixed button actions, never pages or code.
OMB_DISABLE_ANNOUNCEMENTS=1turns it off on a server. - Usage analytics are optional. In the desktop app, the switch in Settings turns them off, and a server's page its window shows starts with them off too (once that server runs the same version); in a browser, each server's page has its own switch. They send counts and fixed words, never keys, names, addresses or content; Usage analytics lists every event.
Your responsibility
- Review approvals, especially shell commands, file writes, and actions in connected services.
- Use dedicated environments for untrusted or destructive work.
- Keep provider accounts and connected apps scoped to what a bot needs.
- Revoke lost companion devices immediately.
- Keep Seira and agent CLIs updated.
Security reports should be filed through the repository's documented security channel rather than disclosed in a public issue before a fix is available.