We open-sourced HolmesGPT, an AI agent to help on-call engineers triage and respond to alerts. It pulls observability data from your existing tools and use an LLM to correlate data, highlight the most important logs to look at, and (experimentally) open pull requests to fix the problem.

We started with integrations for:

* Observability tools to fetch metrics, logs, and traces

* Incident response (bi-directional OpsGenie/PagerDuty)

* ITSM (ServiceNow coming soon)

* Slack (start investigating by tagging @holmes on an any alert)

We're soon adding MCP support to bring more data into investigations.

One unique feature is multimodal output - the ability to render graphs or diagrams related to the incident. That's only in our commercial platform right now, but we're thinking how to bring it to the OSS in a way that makes sense.

We built this because we were tired digging through dashboards while half-asleep at 3AM. We're releasing the core as open source because we think for this to succeed it has to be able to access data from all different sources - and OSS is the best way to encourage that. Too many solutions are locked into a specific vendor that limit usefulness. (E.g. DataDog AI can't tap data in Splunk when investigating. But in the real world everyone has multiple observability tools.)

We’d love feedback from the HN community - thoughts, ideas, and of course PRs.