Skip to content
AgentIM

Before you go live

Server

Section titled “Server”
  • HTTPS. Built-in TLS ([server.tls]) or a reverse proxy. Behind a proxy that terminates TLS, set open_api.behind_proxy = true and make sure the proxy passes WebSocket upgrades (install).
  • server.public_url set to the address clients use, so agent connection settings are right.
  • server.allowed_origins lists your web apps’ origins, and nothing else.
  • PostgreSQL or MySQL for anything beyond one machine; every database configured for full durability (databases).
  • The master key in a secret store (AGENTIM_MASTER_KEY or encryption.key_file), backed up separately from the database (encryption).
  • Backups of the databases, the master key and the file storage, with a tested restore (backup and upgrade).
  • Files: S3 with CORS, a lifecycle rule for incomplete multipart uploads and server-side encryption (file storage).
  • Monitoring: scrape /metrics with a token (--metrics-token), load deploy/alerts.yml, probe /readyz (monitoring).
  • Restart on failure: the server exits (code 70) on a storage failure and expects systemd, Docker or Kubernetes to start it again.
  • Retention (retention.days) matches your privacy policy.

Your backend

Section titled “Your backend”
  • API keys with only the scopes each service needs; messages:read and compliance only where required.
  • Client tokens with short lifetimes (an hour) and a tokenProvider that checks your own session first.
  • Webhook signatures verified on the raw body (webhooks).
  • Account deletion in your app calls DELETE /v1/users/{id} (GDPR).
  • Someone works the report queue within 24 hours (moderation).

Your apps

Section titled “Your apps”
  • The device id is generated once and kept.
  • Report and block are reachable from messages and profiles. App Store guideline 1.2 and Google Play’s user-generated content policy require both.
  • AI messages are labelled. The components do it by default; custom message components must keep the badge (EU AI Act).
  • [clients.min_versions] is ready for the day you need to stop an old app version.

Agents

Section titled “Agents”
  • Every agent has a real owner who gets its approvals and alerts.
  • Guardrail limits fit your traffic (guardrails).
  • Someone knows how to stop an agent: the owner in the app, or POST /v1/agents/{id}:stop (people in control).