AI Agent Olympics submission · Vultr Track · Milan AI Week (May 13–20 2026)

Live cockpit: https://atlas.qubitpage.com
Source repository: https://github.com/qubitpage/vultr-atlas-omniverse

A 3D cockpit for your real Vultr fleet. A spinning globe shows account regions as towers whose height = server count and color = operational health. Click a region to expand real Vultr servers, services, status, maintenance and SLA context. An always-on Gemini copilot ("Atlas") drafts plans, explains tradeoffs, and proposes infrastructure changes — but never executes anything. Every action requires an explicit human click.

This public release also includes the Omniverse/Isaac Sim integration layer used for GPU-hosted digital-twin streaming experiments. The repository has been exported from the live workspace, cleaned of build output and diagnostic scripts, and scanned for literal API keys/secrets before publication.

• Next.js 15 (App Router, React 19 RC, TypeScript)
• Three.js + react-three-fiber + drei — the 3D globe, region towers, pulsing rings, halo, starfield
• Tailwind CSS — dark cockpit palette (teal #5eead4 accent, violet #a78bfa agent accent)
• Zustand — client-side fleet state
• Gemini (@google/generative-ai, model gemini-2.0-flash-exp by default) — streamed via SSE
• Vultr v2 REST API — /regions, /plans, /instances (list / create / destroy / reboot)
• VS Code Web + GitHub Copilot — embedded admin-only editor rail with a chat-only Copilot view and a full IDE view
• NVIDIA Isaac Sim / Omniverse WebRTC — optional GPU-hosted digital-twin streaming assets under omniverse-integration/

cp .env.example .env.local
# Fill in VULTR_API_KEY for real account infrastructure.
# Without it, Atlas shows no account servers/services and asks you to connect the API.
npm install
npm run dev   # http://localhost:3030

• Left: 3D globe. Auto-rotates. Drag to orbit, scroll to zoom. Each region marker = a tower (height ∝ real server count) + a pulsing ring (color = operational health from instance state and Vultr status). City labels billboard above each tower.
• Digital Twin: selected Vultr zone expands into a 3D operational hall containing only real account servers and real account services returned by the Vultr API, plus public Vultr maintenance/SLA context.
• Bottom-left, two cards:
  • Region Panel — clicked region's instances with status, IP, plan, bandwidth and availability data. CPU/load/temperature show N/A unless future guest telemetry supplies them.
  • Recommend — describe a workload (web / api / db / worker / ml / static), set min RAM / vCPU / budget, optionally a region hint. The recommendation engine scores Vultr plans by workload-weighted utility ÷ log(price).
• Right: Atlas Copilot chat sidebar. Streams Gemini responses. When Gemini emits a fenced json proposal block, it renders inline as an action list with an APPROVE button per action. Nothing happens until you click.

Three layers of defense:

1. Model-side: lib/gemini.ts SYSTEM_PROMPT declares Atlas as propose-only.
2. UI-side: components/atlas/AgentSidebar.tsx extracts proposals from streamed text and renders an explicit approval button per action. There is no auto-execute path.
3. API-side: /api/vultr only accepts POSTs with an explicit action field; the chat endpoint is a separate route that has no side-effect tools.

Atlas uses Vultr API v2 for real account inventory when VULTR_API_KEY is set:

• /regions and /plans — datacenter map, plan availability, price, RAM/vCPU/bandwidth class.
• /instances — active cloud servers, OS, plan, region, IP, status/power/server state.
• Best-effort inventory collectors: /load-balancers, /databases, /kubernetes/clusters, /blocks, /object-storage, /firewalls, /reserved-ips, /vpcs, /snapshots, /dns/domains, /container-registry.
• Public status.vultr.com — regional OK / planned service / degraded state and scheduled maintenance windows.
• Public Vultr SLA page — 100% uptime guarantee summary, exclusions and outage-credit table.

Vultr does not expose guest CPU load or physical host temperature for VPS instances through the public account API. Atlas therefore does not invent those values. CPU, load and temperature fields display N/A until a guest telemetry agent or another real metric source is added.

app/
  page.tsx              Landing page
  atlas/page.tsx        Main 3D cockpit
  api/
    vultr/route.ts      GET fleet snapshot · POST create/destroy/reboot
    agent/
      chat/route.ts     SSE Gemini stream
      recommend/route.ts  Plan scoring
components/atlas/
  Globe.tsx             react-three-fiber 3D globe + region markers
  RegionPanel.tsx       Per-region instance list with live bars
  AgentSidebar.tsx      Gemini chat + proposal approval
  RecommendCard.tsx     Workload form + ranked picks
lib/
  vultr.ts              Server-only Vultr API client (real account data only)
  gemini.ts             Gemini system prompt + streaming chat helper
  recommend.ts          Workload-weighted plan scorer
  regions.ts            Vultr region → lat/lon + sphere math
  store.ts              Zustand client store
  types.ts              Shared types
  utils.ts              cn() class merger
omniverse-integration/
  components/OmniverseStream.tsx  WebRTC stream embed component
  server/                       GPU host nginx/docker setup
  akash/                        Akash deployment manifests with placeholder credentials

• The release folder was scanned before push with SECRET_SCAN_FINDINGS=0 for literal tokens, API keys, private keys, and hardcoded passwords.
• .env.example contains empty placeholders only. Real Vultr, Gemini, Speechmatics, NGC, and session secrets must be provided through environment variables.
• Old one-off diagnostic scripts from the private deployment workspace are intentionally excluded because they contained server-specific operational details.

npm run build
npm run start  # listens on PORT 3030

Front with Caddy / Nginx → port 443. The app uses no database; everything is in-memory client state + live calls to the Vultr API on each load.

MIT. Built for the AI Agent Olympics.