Local AI integration — fleet telemetry to local LLM

[castrojo]

Contributors run a local LLM that reads fleet telemetry from Loki and Prometheus and surfaces recommendations: performance issues, anomalies, suggested fixes. The Star Trek computer use case — your homelab explains itself to you, privately, on your own hardware.

CNCF context: OTel Collector ships telemetry. Loki (Grafana Labs OSS) stores logs. Prometheus (graduated) stores metrics. The LLM reads from these via MCP tools (linux-mcp-server, goose) or direct API. No external AI services required.

Principle: Local first. BYOLLM. No telemetry leaves the contributor's network. Frontier models supported but local models are the target.

[castrojo]

"You don't need to be a nation to have digital sovereignty. You just need the same technology they use."

[castrojo]

Ok some feedback from jeefy and mrbobbytables. The whizbang management is different from the installer. Idea:

• projectbluefin/knuckle - installer, no opinion, just install flatcar
• projebluefin/bluespeed - uses knuckle for a kickass homelab experience.

Then we can call all the management, self hosting, AI, magical dashboard cluster thing "Bluespeed" as just a collection of local-AI first driven private networks or somesuch.