Purchases and Costs

Orca Connect: attach your own GPU to Claude Code

One outbound connection, spillover off by default

Orca Connect publishes your home GPU¹ into the OrcaRouter catalogue as local/your-rig/your-model. It opens one outbound connection with no port forwarding, and a local-only key blocks hosted billing. Prompts² still pass through the router, and the free tier is a current condition, not a promise.

What actually changes

Run the connector on your machine and it finds the local runtimes, then lets you pick which models to share. They appear in the catalogue as local/your-rig/your-model.

Callers keep using the key and endpoint they already have; from a laptop or a phone the request still reaches your machine through the router.

There is no port forwarding, no tunnel to configure and no new firewall rule, and nothing is exposed until you pick a model to share.

The connector only shares models already loaded in that machine's memory, so a 24 GB box lists whatever fits inside 24 GB.

A desk where an agent laptop connects to a local inference server on a separate device — The laptop only sends requests; the model runs on the machine under the desk, and that machine opens the connection outward.

Spillover starts switched off

With spillover on, a hosted model answers only for the reasons you allowed, only the model you picked, and only within the monthly cap. With it off, none of that happens.

Response headers record where the work ran: X-Orca-Billed: false for your hardware, and a reason plus expected cost when a hosted model took over.

What spillover records
Situation	Response	Billing
Your machine answered	X-Orca-Billed: false	KRW 0
Machine off, spillover off	503 local_connector:unavailable	None
Machine off, spillover on	X-Orca-Spillover-Reason: rig_offline	Hosted rate
Monthly cap reached	Refusal	None

Point a coding agent at your machine

Running orca-connect claude starts Claude Code against your model for that session only, leaving the agent's own configuration alone.

Swap claude for codex, opencode, gemini, qwen, kimi, copilot, aider, goose or pi and the matching CLI³ sees your model the same way.

Which CLI each command launches
Command	CLI it launches	When to use
orca-connect claude	Claude Code	You already use Claude Code
orca-connect codex	Codex CLI	Long terminal sessions
orca-connect aider	Aider	File-by-file edits
orca-connect goose	Goose	Tool-heavy workflows

Prompts still cross the router

A key marked local-only keeps inference⁴ off hosted models and blocks billing, but the request itself still travels through OrcaRouter.

The distinction matters because where inference runs and where data travels are different questions; if you handle page content or internal documents, settle this first.

An always-on server corner with airflow, wired network, backup power, and a maintenance notebook — A machine you call from outside has to stay awake; settle sleep and reboot policy first.

Install and share once

Install the connector and approve the machine once in a browser; a headless server approves with a code instead.

orca-connect status lists the runtimes it found, and the console is where you pick which to share; the tray or orca-connect expose handles the first pick.

Check status

orca-connect status
orca-connect expose

A rack of cable-connected server equipment beside a laptop — The point is not opening a port: no inbound means no extra firewall rule.

Where it falls short

It is a commercial service. It is free today, but pricing and terms can change, so check them before putting it into a workflow.

The machine has to be awake with a live connection; sleep or a reboot makes outside calls fail.

For same-room use a plain LAN server is faster and simpler, so this connector is not a LAN replacement; it pays off when you call from outside.

The licence of a shared model is your responsibility; check the commercial terms on its model card.

Terminology notes

GPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the text
Prompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
CLI — Short for Command-Line Interface: operating a program by entering commands in a terminal.
Back to the text
Inference — The process of using a trained model to compute an output for an input. Here, local inference means running the model on the user’s device.
Back to the text

Frequently asked questions

What should you check first in this article?: Orca Connect publishes your home GPU into the OrcaRouter catalogue as local/your-rig/your-model. It opens one outbound connection with no port forwarding, and a local-only key blocks hosted billing. Prompts still pass through the router, and the free tier is a current condition, not a promise.
How much memory does it need?: The connector only shares models already loaded in that machine's memory, so a 24 GB box lists whatever fits inside 24 GB.
When is it better not to choose this?: For same-room use a plain LAN server is faster and simpler, so this connector is not a LAN replacement; it pays off when you call from outside.
Where can you check the conditions and sources?: The licence of a shared model is your responsibility; check the commercial terms on its model card.