One interactive map of the real system: the 9-node compute fleet, the request path from
magnaxis.ai down to individual engines, the 508 engines live in the dispatch catalog
(out of 3,520 engine directories in the source tree), and the data/AI layer underneath. Every number on
this page was pulled from a live probe or a config file on the fleet — not from a spec doc. Click any card to drill in.
The path a customer request actually travels, node by node. Click a layer to see the real ports, processes, and evidence behind it.
Every node's role, verified this session via ssh -F /root/.ssh/aria_config <node>
(hostname-pinned — never resolved by ambient DNS) plus the live :8795 catalog map and node2's own
fleet blueprint. node5 = prod edge only, internals intentionally not probed
The dispatch-router (:8795 on node2) is the single gateway every top-level engine
call goes through. It fronts 508 engines spread across 5 physical backends — confirmed by curling its own
/health and /catalog endpoints live.
/catalog → map):
All 508 engines currently registered in the live :8795 catalog, grouped into 31
real capability domains by name/function (this session's classification — the router itself does not tag
domains). This is one layer above the 3,520 total engine directories on node2's source tree
(engine-master-spec/engines/) — most of the difference is sibling/leaf engines that top-level
orchestrators compose in-process (loadSib) rather than separately network-dispatchable members.
Click a domain to see its members.
What actually holds state, and the single AI provider everything routes through.
The end-to-end chat→delivery path, with the honest current state of each hop (verified live today — flags can and do flip).
The gates every engine and every delivered surface has to pass through — anti-fabrication first.