Run a chat through OpenCode on your PC

If you use OpenCode, QuorX can now run a chat through it: every provider you connected in OpenCode (your plans, keys and sign-ins) shows up in QuorX's model picker. A chat's main model runs there, and every tool, file, command, browser step, check and schedule works the same as on any other model. Everything here is in **⚙ Settings → Models → Apps on this PC**.

Switch it on

QuorX finds OpenCode on this computer and the providers you already connected in it. Tick OpenCode on this PC and press Test the connection: it shows how many models OpenCode offers and from which providers. Nothing to pay, and QuorX holds no key for it. If OpenCode isn't installed yet, the page shows the command: npm i -g opencode-ai (in Windows itself, not only in WSL: QuorX looks for it on the Windows PATH). To add a provider, run opencode auth login in a terminal, then test the connection again.

OpenCode's own free models (its OpenCode provider, at no price) answer only inside OpenCode's own app. They show in QuorX's picker but locked: hover or click says why, and a call to one is refused before it goes out.

Once it's on, its models show in the chat model pickers, tagged This PC. On one test computer that was 854 models. The Where it runs switch at the top of the picker shows one source at a time, so a long list stays short.

How it works

QuorX starts OpenCode's own server in the background when a chat needs it, and stops it after a quiet quarter of an hour. OpenCode makes each call itself, so your plans and sign-ins work as they do in OpenCode.

OpenCode runs with its own tools off. QuorX describes its tools to the model, and the model writes its calls as text that QuorX reads back (in QuorX's format, or the model's own, such as DeepSeek's). Each call is a fresh OpenCode session that QuorX deletes afterwards, so your own session list stays as it was.

What runs where

The chat's main model runs through the app (its summaries of a long chat too). Helpers, the reviewer, the prompt enhancer and fresh eyes run on OpenRouter by default, and each can be set to a model tagged This PC in its model picker. For the reviewer, pick one that sees pictures: it looks at screenshots. Deciders and paid media always run on OpenRouter. You can switch a chat between OpenRouter and an app on your PC at any time. The conversation carries over.

What it costs

When a provider reports a call's cost, QuorX counts it in the reply, the project and your daily caps; otherwise it counts the call's tokens at the provider's list price. Plan and free models count $0, and each step on a plan shows what it counted against the plan, at list price.

No OpenRouter key yet?

You can still chat: with no OpenRouter key, the models tagged This PC run, and the others show locked until you add a key. The deciders and paid media need OpenRouter, so they are off until then. Helpers, the reviewer, the prompt enhancer and fresh eyes are off too, unless you set them to a model tagged This PC: then they keep working. The chat says which ones are off.

When it fails

QuorX never switches your model behind your back.

• The reply stops and says which model failed and why (an expired key, a used-up plan, a free tier that ran out). When a setting in your account there blocks the model (OpenCode Go: "This Go model requires Global regions"), it says so at once: change it in your OpenCode workspace (Privacy → Global), or pick another model. The work done so far stays. Pick another model in the chat's model button and send a message to carry on: the conversation carries over.
• A model that failed is skipped by every chat, this one included, for a while instead of each one waiting on it: 10 minutes after a refused key, up to 30 minutes after a used-up plan, up to 6 hours after a free tier says no, and under 5 minutes after a server error. A passing Test the connection tries them all again at once. A short overload or rate limit is waited out first, the way OpenCode retries it. Each provider is separate, so one model reaching its limit leaves the others working. Only the app itself missing or unreachable affects them all.

Any text streamed before a failure moves into the steps, so the reply isn't written twice. No tool call from a failed answer ever runs.

If you'd rather have the reply carry on by itself, choose a fallback in **⚙ Settings → Models → Apps on this PC → When a call fails: **Same model on OpenRouter (for example opencode-local/opencode-go/glm-5.3 becomes z-ai/glm-5.3) or Always this model. A step that falls back shows ⤵ with why, and costs OpenRouter's price.