finds.dev

VersusControl/versus-incident

The self-hosted AI SRE agent. You don't have to monitor your systems alone. Versus Incident will keep track of your services, applications, and systems, detect new or unexpected problems, and provide your team with the evidence needed to investigate.

View on GitHubHomepage ↗

★ 908GoMITupdated Oct 2026

// the take

A self-hosted incident management gateway in Go that does two jobs: accept webhooks from Alertmanager/Grafana/Sentry/CloudWatch and fan them out to Slack, Teams, Telegram, Viber, email and PagerDuty/Opsgenie-style on-call escalation, and separately run an 'AI SRE agent' that reads your logs, clusters patterns, and pages you only on genuinely new error signatures. Good fit for a small ops team that wants PagerDuty-adjacent routing without the SaaS bill, and is willing to tune pattern detection themselves.

One webhook endpoint understands Alertmanager, Grafana, Sentry and CloudWatch SNS payloads out of the box with a shared default template, so you're not writing N parsers for N tools. The agent's training → shadow → detect progression is a sane way to earn trust in anomaly detection before it can actually page anyone, rather than wiring an LLM straight to PagerDuty on day one. Tool integration surface for the agent (k8s, Elasticsearch, Loki, Prometheus, Tempo, Splunk, Graylog, CloudWatch) is unusually broad for a project this size. Ships a real Helm chart plus an HA example with PodDisruptionBudget and NetworkPolicy, so someone has thought past `docker run`.

The pattern catalog and incident history only persist to local files today — Redis and database backends are documented as stubs — so the HA example's premise (multiple replicas, shared state) doesn't actually hold for the agent's core data. All admin and agent API access rides on a single shared `gateway_secret` exchanged for a cookie; there's no per-operator identity or RBAC, which is a gap for any team bigger than one person. Anomaly detection quality depends entirely on your `similarity_threshold`/regex pre-filter tuning, and `auto_promote_after` silently reclassifies a pattern as 'known' after 100 sightings — a real recurring problem could stop alerting with no signal that it happened. At 908 stars this is young for software whose entire job is deciding whether to wake someone up at 3am; expect to find edge cases PagerDuty/Opsgenie worked out years ago.

We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
908
Forks
166
Language
Go
License
MIT
Last commit
21 Oct 2026

// nearby

If versus-incident isn't quite it.

All repos →