Your GPU joins the router. The cloud only takes over when you say so.
Orca Connect dials out from the machine that runs your models. They join 200+ hosted ones behind your usual key, and OrcaRouter charges nothing for them.
- Outbound only: no port forwarding, no tunnel to configure, no firewall rule. Nothing is reachable until you pick the models to share.
- Spillover is off until you turn it on. Then a request moves to a hosted model only for the reasons you allowed, to the models you named, under a monthly ceiling.
- A response your hardware served carries X-Orca-Billed: false; one that spilled to a hosted model carries the reason and its estimated cost.
Share a rig, not a password
Send someone a link to the models you choose on your machine, with an expiry and a request limit, and revoke it whenever you like. They sign in with their own account, so your key never leaves your hands.
Keys that never reach a hosted model
Mark a key local-only and its requests are only ever served by a local rig: never by a hosted model, and never billed. They still pass through OrcaRouter on their way, so this decides where inference runs, not where a prompt travels.
Your coding agent, pointed at your rig
Run orca-connect claude and Claude Code starts against your own models, for that session only. Swap claude for codex, opencode, gemini, qwen, kimi, copilot, aider, goose or pi to do the same for Codex CLI, OpenCode, Gemini CLI, Qwen Code, Kimi Code, Copilot CLI, Aider, Goose and pi. The agent’s own settings are left as they were.
The rig under your desk, reachable from anywhere.
Install it, approve the machine once in your browser (or with a code on a headless server), and tick the models to share. The connector opens one outbound connection and your models appear in the catalog as local/your-rig/your-model. Call them from a laptop, a phone or a cloud agent with the key you already use. Those calls pass through OrcaRouter on their way to your machine; a program on the machine itself can still talk to the runtime directly.
Free until you say otherwise.
With spillover off, close the laptop lid and callers get a clear refusal, not a bill. Turn spillover on for the reasons you care about (rig offline, rig busy, prompt too long, a runtime error, a request that needs tools), pick which hosted models may take over and set a ceiling. Or go the other way: keep writing gpt-4o-mini in your code and let your own GPU answer whenever it runs the model you mapped to that name; when it can’t, gpt-4o-mini answers as it always did and is billed as usual.
How it works
- Sign in once — Approve the machine in your browser, or with a code on a server that has no screen. The connector saves the key it receives in its own config file on this machine; nothing is pasted into a terminal.
- Pick what to share — The connector finds the runtimes on this machine by the ports they listen on, and tells them apart by asking each something only it answers; orca-connect status lists what it found. You choose the models to share in the console; before the machine first connects, the tray or orca-connect expose can make that first choice. OrcaRouter sees the list so it can offer you the choice, and sends requests only to the models you share.
- Go online, outbound only — One connection, opened from your machine, carries every request as a stream of its own. Your router never sees an inbound connection, and your models are listed as local/your-rig/your-model.
- A request reaches your rig — A coding agent, a phone or someone you shared a link with calls the usual endpoint with a key. OrcaRouter checks that key’s policy, finds the connection your machine holds and sends the request down it. Your machine passes it on only if the model is one you shared.
- The answer says it was free — Tokens stream back the same way. The response carries X-Orca-Billed: false, and the request is logged at $0.
- The lid closes — Once OrcaRouter sees the connection is gone, a caller with spillover off gets 503 local_connector:unavailable: a clear refusal, and nothing is billed. With spillover on for an offline rig, a hosted model you allowed answers instead, under your monthly ceiling; that answer is billed and carries X-Orca-Spillover-Reason: rig_offline.