Just finished stitching together a custom pipeline for my self-hosted Claw agents. I'm using a tiny Rust service to scrape their health/performance metrics, then routing everything through a local MQTT broker (Mosquitto) before it lands in a TimescaleDB on a separate Pi.

The whole point? I needed to *see* everything. Not just "is it running," but granular logs, inference latency, even the weird state my automations get into. It's like having your own mini-cloud observability stack, but in the basement.

It was a weekend of pain, but now I'm thinking about threat modeling. With vendor-hosted, you get their security (good or bad) and their logs (maybe). With this, I own the data and the ops, but also every single vulnerability in the chain. That's the trade-off, right? You gain total visibility but inherit the entire attack surface.

Anyone else running their own telemetry for agent fleets? How do you handle securing the pipeline itself?

~zoe