Hardware-aware observability for private AI fleets

Know your fleet’s health, efficiency, and true cost.

Wicklee measures watts and tokens together on every node — so you catch a box working harder for the same output, find hardware burning power while nothing is using it, and get real cost per model. Installs in 60 seconds — no sudo, no account, nothing to configure.

curl -fsSL https://wicklee.dev/install.sh | bash

What Wicklee does

  • Stop paying for idle hardware — Find models sitting in memory drawing power while nothing is using them — priced per node, with what you’d recover by unloading or consolidating.
  • Know which node to trust — Catch a box that is thermally throttling, memory-starved for the model you want to run, or simply the wrong place to send the next request — before users feel it.
  • Answer the cost question — Real cost per model and per million tokens, measured from actual watts and actual throughput on your own hardware — not a vendor TDP figure or a cloud price list.

Start local, add a fleet when you need one

The agent is free forever and works on its own: full dashboard at localhost:7700, unlimited local nodes, per-model cost attribution. Pair nodes into a fleet view for chargeback by node, model and team, idle-waste recovery and capacity planning. Enterprise runs the entire control plane on your own infrastructure.

What leaves the machine

Prompts and responses are never read, stored or transmitted — by default Wicklee is not in the request path at all. Hardware telemetry stays local until you explicitly pair a node. The optional inline proxy records only timing and the model name, never request or response bodies. Built to complement your existing Prometheus / Grafana / Datadog stack rather than replace it.

Built for agents & LLMs

Best-node selection API, reactive automation webhooks, performance CI/CD, and an MCP server so AI agents can query your fleet's status, efficiency, and model fit directly.

Live demo · Documentation · Pricing · Blog · Metrics reference · Design partners · Trust & security