Close the cloud door.
Choose Local only. Remote candidates become ineligible. Then switch your local model offline: the request stops instead of falling back to a remote provider.
A clear view of every decision. Set a policy, follow the selected path, and compare three sample providers.
Weigh estimated cost, latency, and reliability together.
Switch it off to test an unavailable route.
Best combined sample score: 40% estimated cost, 40% latency, and 20% reputation.
| Provider | Est. cost | Latency | Reliability* |
|---|---|---|---|
Trusted peerSelected | Est. cost$0.0010 | Latency410 ms | Reliability*98.8% |
Configured APIEligible | Est. cost$0.0032 | Latency240 ms | Reliability*99.5% |
Your local modelEligible | Est. cost$0.0000 | Latency840 ms | Reliability*96% |
*Sample reliability is a reputation proxy. Bars compare sample latency against 840 ms.
This preview uses illustrative scoring. It does not contact your router or send an inference request.
Choose Local only. Remote candidates become ineligible. Then switch your local model offline: the request stops instead of falling back to a remote provider.
Prices are synthetic request estimates. Latency is sample time to response headers. The demo uses reliability as a simple stand-in for reputation.
These providers share the sample alias local-chat. In Cairn, you choose which models can substitute for one another. A shared alias does not promise equal quality.
Balanced uses an illustrative normalized score: 40% estimated cost, 40% latency, and 20% reputation. Lowest cost or Lowest latency ranks eligible providers by that single criterion. Local only filters out remote providers before selection.
All numbers and identities are sample fixtures. This page makes no provider calls and does not connect to your local router. The actual router uses configured prices, estimated token usage, operator priors, and locally observed reliability. The peer price illustrates a future paid-capacity scenario; public-chain payments remain disabled.
Start with the synthetic beta demo, then connect your own models.