Autopilot continuously tunes every pod, node, and autoscaler to real usage — in place, inside your cluster. Zero downtime, flat cost, no telemetry.
Built on the primitives Kubernetes gives you — not around them.
The gap between what your cluster reserves and what it actually uses is money — most fleets run far below what they pay for. Right-sizing reclaims that difference, and Autopilot reclaims it continuously, recomputed as your usage shifts.
Four systems, one loop:
KEP-1287 in-place resizing matches every workload to its real CPU and memory usage, recomputed continuously and applied with no restarts, no rescheduling, no dropped connections.
When pods can't fit, Autopilot picks and provisions the right machine per pod instead of a one-size pool — spot and on-demand together — and drains the empties when demand drops.
Analyzes every HPA and only touches the ~27% that are misconfigured: oscillation stabilized, min replicas right-sized, elasticity intact.
Bin-packs fragmented capacity onto fewer, cheaper nodes, with a proven placement simulation and PDB-aware draining before a single pod moves.
The same tuning that cuts the bill keeps production standing: workloads floored so they don't fall over on their busy day, autoscaler flapping damped before it cascades, consolidation that never strands a pod, and resizes with no restart — so optimization never becomes an outage.
When right-sizing is automated and provably right, resource management stops being a team's job. The cluster stays correct without anyone standing over it — your DevOps team covers more ground, not more tickets.
The reason autonomous optimizers don't get turned on is fear they'll break production. Autopilot is built the opposite way:
Savings-based tools charge you more precisely when they work best. Autopilot is priced by footprint, not savings: a small, fixed monthly cost that doesn't grow with your cluster or your results. The better it works, the more you keep — 100% of it.
And beyond the cluster itself: observability for the managed services around it (databases, caches, network) and per-package cost attribution. The same managed cluster runs the Docker microservices you bring — including self-hosted LLMs via Ollama — right-sized by the same loop, inside your infrastructure boundary.
The docs cover everything: a 15-minute install, the full optimization loop, scheduled scaling, resource overrides, and pod tools. → autopilot.docs.xano.com