Route a request
Type anything, or pick a sample. The rules decide first, exactly as the router's Code nodes do; the pill on the right shows the tier and the rule that fired.
Attach a document (the real router reads it separately; size rules count both)
GPU busy: the real gate reads a state row the desktop writes when a game starts; a heartbeat older than 3 minutes counts as not busy.
short-plain
The rules did not decide. In the real hub a 4B local classifier labels only this middle (its allowed answers: local-fast, local-think, frontier, design, human, big node; it is told the tier definitions and nothing else) and the document guard is applied again to its answer. No model runs in this demo, so it stops here.
Why
Size
What runs next
See this branch in the router graph
Ported line for line from the router's Code nodes (Rules, GPU-busy gate, Merge Decision, big-node async check), export of October 2026, verified against the original on 152 synthetic cases. The router's own evaluation on 50 sample requests: the rules decide 45 of 50 and get 29 right; see the project page. Two names were changed for publication: the big-memory tier is called "big node" here, and the two laptop nodes are named by their memory and GPU.
Explore the workflows
29 workflows, 539 nodes, 591 connections shown; personal and home workflows, the voice and chat relays and the spend controls of the same instance are not included.
main connection model / memory / tool feeding an agent node carries a credential (count only) disabled node
Who can call what
10 MCP endpoints across 5 tool groups shown; a sixth group is left out. Pick a group to see its tools.