Fleet Explorer: Interactive Topology

The June fleet snapshot used a hand-coded SVG. It reads well, but it's a wall — fixed size, fixed zoom, and a pain to edit when the fleet shifts. This is the same topology as a living canvas: pan, zoom, and read the detail at whatever scale makes sense. Built with tldraw, authored programmatically from a JSON spec, embedded read-only.

Drag to pan · scroll or pinch to zoom · read-only · made with tldraw

Why interactive

Static diagrams force a tradeoff: zoom out for the whole picture and the per-node detail gets tiny, or zoom in for readable text and lose the topology. The fleet has enough nodes now — six boxes, two inference engines, a TP=2 cluster bond, cloud fallback paths — that the tradeoff was starting to hurt. A canvas removes it: the overview is the default view, and the fine print (ports, throughput, roles) is a scroll away.

The pipeline

The diagram isn't drawn by hand. It comes from a structured spec — nodes, labels, edges, colors as data — which means:

• Fleet changes are edits to JSON, not redraws. A new box or a port change is a one-line diff.
• The same spec renders three ways: this interactive embed, a static SVG export (for RSS readers and archival), and a round-trippable .tldr file that opens in the real tldraw editor for manual touch-ups.
• Text stays exact. No AI-generated image artifacts, no hand-aligned SVG text nodes. Labels render from the spec, crisp at every zoom level.

documents/fleet-spec.json     # source of truth (git-tracked)
        ↓  Vite + React build (dark theme, license baked)
build/dist/                   # static embed bundle
        ↓  scp to droplet
/tldraw-embeds/fleet-explorer/  # served by nginx, iframed here
        ↓  Puppeteer + editor API
exports/fleet.svg / .tldr     # static + editable artifacts

What you're looking at

The topology itself is unchanged from the June snapshot — M3 Ultra dedicated to Kimi K2.6, M5 Max running the embed/reranker stack, the two DGX Sparks bonded over QSFP 200G serving DeepSeek V4 Flash in TP=2, Forge as the agent hub, and the Mac Studio routing production traffic without hosting a local model. The reference post has the full per-machine breakdown.

Cost notes

The interactive embed is heavier than a static SVG — the editor chunk lazy-loads at ~526 KB gzipped (the page's initial payload stays around 61 KB; the canvas only loads when the iframe mounts). For a dedicated explorer page that's the right trade. Everyday posts keep static SVG exports, which weigh nothing. The canvas is read-only: pan and zoom work, editing doesn't, and the whole thing is served from the same droplet — no third-party calls.

What's next

• Live status on the nodes. The spec is data; wiring it to real health checks (which endpoints are up, current throughput) turns the map into a dashboard.
• Agent trace maps. The same pipeline can visualize a Hermes delegation run — tool calls, handoffs, failures — as a spatial graph instead of a log dump.
• Static fallback in every post. Any interactive embed ships with its SVG export alongside, so nothing depends on JavaScript to be readable.

Reference: Local LLM Fleet — June 2026 · This is a living document — as the fleet evolves, this post gets updated.