Steer the engines while they run.
Gimbal is the operations console for the AccelMars platform — live deep health for every engine, routing and cost levers, and the incident drill that contains a failure in seconds. Deployable is the demo. Operable is the product.
In the constellation
Gimbal highlighted in the live map — hover or tap any node to explore.
Deployable ≠ operable.
Your AI platform is live. Then an engine starts fabricating answers, or a model call hangs, or the cost curve bends the wrong way — and every question dead-ends in a terminal. Which engine? Since when? What's it costing? How do you stop one component without redeploying everything?
A platform whose every lever is build-time has no operations story — only rebuilds. The fix isn't another dashboard of charts; it's a control surface: the levers the platform already exposes, rendered where an operator can see the fleet and act in seconds.
Every serious runtime grows exactly this organ — the JVM has Mission Control, the Linux server has its cockpit, every cloud has its console. Gimbal is that organ for the AccelMars engine fleet.
Four verbs, one console.
Monitor
the fleet, breathing
Deep health per engine — error rate, latency, request flow, spend — ticking live on one board. Not a shallow 200; the health that tells the truth.
Configure
levers, not redeploys
Quality-tier routing, limits, and keys applied to the running system. Rebind a tier and see the cost-per-request change before you commit.
Intervene
the incident drill
Shed a misbehaving engine in one guarded click — 503 at the edge, zero blast radius. Re-enable and the recovery curve is watched until nominal is confirmed.
Audit
the trail that writes itself
Every lever, proposal, and verdict — attributed, timestamped, immutable. The console that operates the fleet appears in the trail it renders.
Don't take our word for it. Pull the kill-switch.
Watch the fleet breathe
Five engines with live error rates, latency, and spend. One of them — Crucible — is about to have a very bad morning. The system flags it the moment deep health degrades.
Shed the failing engine
Pull the kill-switch. Crucible answers 503 at the edge instantly; the other four engines never feel it. Reversible, and audited under your name.
Bring it back, audited
Re-enable and watch the recovery curve decay until the system — not you — confirms nominal. Then read the audit trail: the whole incident, attributed.
Fleet nominal · 5/5 · watching deep health…
standard→deepseek/deepseek-chat · 212 rpm · $0.0004/req
summary lane nominal · p95 312ms
quick→gemini/gemini-flash · p95 96ms
This is one screen. The full prototype is a control room.
The embedded demo shows the drill only. The full clickable prototype adds the complete fleet board with streaming traces, per-engine deep health, tier-to-provider routing with live cost previews, the full audit trail — and the role switch between Operator (full levers) and On-call AI (propose-only, fails closed). Your session runs entirely in the browser.
Open the Full PrototypeA console you can trust with the keys.
Thin renderer
Gimbal owns no runtime truth and computes no control decisions. Every reading comes from the platform control plane; every action round-trips through it. A missing lever is a platform gap to file — never app-side logic.
Fails closed
Admin auth on every lever, no anonymous read tier, no degraded-auth mode. On-call AI sessions hold propose-only authority: they stage, the operator approves, the trail records both.
Provider neutrality
Provider and model names render as data from gateway state. None exist in Gimbal source — the same invariant that governs the gateway itself, kept all the way up to the pixels.
Everything audited
There is no unaudited path through the control plane — kill-switches, rebinds, proposals, approvals, even the system’s own anomaly flags land in one immutable, attributed trail.
Three surfaces, three protagonists.
Run the platform like a flight, not a fire.
Start free with three teammates on Atlas today, or dive into the operating system that powers them.