Paeon¶
The on-call colleague for your self-hosted server. Alert triage, read-only investigation and approved remediation, with a hard line between what a model may think and what code may do.
It is 03:12. Your phone buzzes: SLO fast burn. Is it the nightly update again, or is the disk
really full this time? Paeon sits next to your alerting and answers that question before you are
awake enough to ask it.
Out of the box Paeon takes no action on your server
It records, correlates, triages, investigates and drafts postmortems. Acting is a second process you add on purpose, with a closed catalog of actions, and nothing runs without your Go until you say otherwise. It is alpha software, and the acting part has not yet run on a real server for long: run it next to your alerting, never instead of it.
Start here¶
| Quickstart | One machine, no model key, six steps: an alert in, an incident out, the console open. |
| Prometheus quickstart | A compose project you can throw away: break a service, read the investigation. |
| How it works | The pipeline, the stages, and why the part that thinks cannot act. |
| Videos | Paeon explained, and one incident followed from the alert to the fix. |
Why Paeon¶
- The model proposes, code disposes. No shell, no free-form command, no write credential for a model, in any stage. A plan is a list of actions from a closed catalog, checked again by a separate executor that trusts nothing the model touched.
- Go / No-go on your phone. A signed, single-use, expiring button, bound to the exact plan. Opening the link never applies anything.
- Autonomy is earned, per action and per container. Paeon proposes a promotion after a record of verified runs; you make it, in git. One failure, one Undo or one "This was wrong" takes it back.
- Investigation with evidence. Every claim in an analysis is tied to the tool call that supports it. A claim that cites no real call is marked as unverified.
- Allow-lists first, redaction second. A log stream, a metric or a container that you did not list cannot be read, and what is read is scrubbed of secret-shaped strings before it leaves.
- Everything is an event. One append-only SQLite log: every alert, decision, model call (with its cost) and tool result.
- Bounded by design. Budgets per incident and per day, a storm guard per hour, limits on tool calls and time. A spent budget never hides an alert: Paeon fails open.
- Your models, your gateway. The Anthropic Messages API and the OpenAI-compatible chat API, each with a configurable base URL. Local servers work.
- Small. One container, one SQLite file, one YAML file. No vector database, no queue.
Paeon is an independent project, not affiliated with or endorsed by OpenObserve, uptime-kuma, Prometheus, Docker, ntfy, Anthropic or OpenAI; those names belong to their owners.