Command line, API and MCP¶
Command line¶
| Command | |
|---|---|
paeon serve |
Run the core |
paeon executor |
Run the executor (only with actions.enabled) |
paeon incidents |
Incidents as JSON lines |
paeon events --kind alert.firing |
The event log as JSON lines |
paeon digest --hours 24 |
The digest of the last hours |
paeon feedback <id> up\|down |
Say whether triage was right |
paeon autonomy |
Each action's level, its track record, and the promotions it has earned |
paeon stop / paeon resume |
The kill switch, next to the executor |
paeon reinstate <alert> <action> <container> |
Lift a demotion, next to the executor |
paeon notify <channel> |
Send a test message through one notification channel |
paeon source <name> |
Say what one source gives, and ask its service one question |
paeon mcp |
Serve the record read-only to an assistant, over MCP on stdio |
paeon replay triage\|investigation |
Re-run recorded incidents against the model and prompts of today |
The same is available over HTTP behind the bearer token: GET /incidents, GET /incidents/{id},
POST /incidents/{id}/feedback, GET /digest, GET /postmortems, GET /autonomy, and
GET /brief (what is open, what waits for your answer, what waits for a review: made for a
daily brief).
Alerts come in by POST /ingest/<name>, one address per source, and planned work
by POST /ingest/change.
Ask it from your assistant¶
paeon mcp is a read-only MCP server over the record:
incidents, their findings, postmortems and runbooks. It opens the database file directly, so it
needs no network path to the service. In Claude Code, from the repository of your server:
Then "has this happened before?" is a question it can answer. What it returns is redacted like everything a model reads, and it has no tool that writes.
Before you change a prompt or a model¶
Every triage call recorded what it was given, and every investigation its tools and what each
one returned. paeon replay sends the same again to the model as it is configured now:
It compares with what was said then and with what you said about it. A verdict you rated right
that changes is a regression, and one that would now be called noise is reported loudest; the
command exits 1. For investigations the old and new root causes are printed side by side, and
a call the model did not make at the time answers "not recorded". --export corpus.jsonl writes
the cases to a file you can keep in git, --corpus replays from it, --dry-run lists them.
A replay writes nothing to the event log and counts against no budget, but it does call the model: it prints what that cost. Pin the model for it; with a gateway choosing, a prompt change and a routing change cannot be told apart.