Pheme is an intelligence layer for Prometheus and Grafana stacks. It sits alongside your existing observability tooling, watching the same metric streams, and surfaces anomalies before threshold-based alerts fire.
| Capability | Description |
|---|---|
| Anomaly detection | Passively watches Prometheus remote_write streams and flags deviations before alert rules trigger |
| Signal enrichment | Publishes annotated alerts back to Alertmanager and Grafana without replacing them |
| Incident summariser | Generates AI-written incident summaries for operators when anomalies cluster |
| Runbook RAG | Retrieves relevant runbook sections and past incident context at alert time |
| Pluggable LLM backend | Bring your own model: OpenAI, Anthropic, Bedrock, or local inference (Ollama) |
Pheme is purely additive. Remove it and your existing stack is unchanged.
- Not a monitoring system. Prometheus remains that.
- Not a metrics database. Prometheus and Thanos remain that.
- Not a Grafana replacement. Grafana remains that.
- Not a replacement for alert rules. Those keep working alongside it.
Apache 2.0.
Status: early development. The architecture and v1 are actively being built.
The capabilities below describe the project's goal, not a finished release.
Not production-ready.
Inspired by the mythological Pheme: the first to know, the one who never slept.