Reference
Limits and what is not built yet
An honest list of caps, timeouts and missing features, so nothing is a surprise.
Everything here is current as of the beta. Where something is missing, the page that covers it says so too.
Caps and timeouts
| Thing | Limit |
|---|---|
| Uptime monitors | 1 on Free; 50 per project on paid plans |
| Uptime intervals | 1, 5 or 10 minutes |
| Downtime threshold | 3 consecutive failed checks |
| Server notes | 2,000 characters |
| SSH session idle timeout | 30 minutes |
| SSH connections per session | 5 |
| Assistant command timeout | 20 minutes |
| Approval card timeout | 10 minutes (counts as No) |
| S3 upload through the app | 50 MB per file |
| Team logins | 6 on the Team plan |
| Activity log retention | 365 days |
| Sign-in session | 30 days |
| AI credit on a new account | $2 (Free) |
| Projects | 1 on Free |
| Servers | 2 on Free |
| Managed databases | 1 on Free |
| File storage | 100 MB on Free, 1 GB on paid plans |
See Plans for the full Free / paid split.
Not built yet
Alerts. No email, SMS or webhook notifications from uptime, from health checks or from CloudWatch alarms. Uptime records outages for you to look at. Keep your existing alerting.
Google Cloud. Shown as Coming soon in the project type picker and not selectable.
Billing. No card, no checkout. Everything is free during the beta.
Background health checks for EC2 instances. The 15-minute sweep and the health cards cover Server Only projects. EC2 instances can be opened and used; they do not get automatic health cards.
Activity log export. Readable on screen; no CSV or PDF export yet.
A screen for adding MySQL connections. The capability exists; the form does not. See MySQL.
Email password reset for team members. Their admin resets it for them.
CloudWatch alarm management. Alarms are listed read-only; create and edit them in the AWS console.
IAM user management. Read-only by design — create and change IAM users in the AWS console.
Behaviour worth knowing
Restarts end sessions. When DevOps Agent is updated, open SSH sessions close and waiting approval cards are cancelled. Nothing runs that you did not approve; you simply reconnect.
Long chats get expensive. A chat carries its whole history into every reply. Start a new chat per problem.
Stopping a running answer is imperfect. Pressing Stop stops the browser listening; a command already running may still finish. The activity log is the truth about what ran.
Reasoning effort only applies to OpenAI-hosted models. On other providers the setting is shown but ignored.
Attached servers are per project, not per chat. Unlocking a server for one chat makes it reachable from other chats in the same project.
Where limitations are discussed in detail
- The safety switch — what Read mode cannot promise
- How we protect your credentials — what the model can see
- Activity log — what the log cannot see
- Roles — where role boundaries are guardrails rather than proofs