Your own AI apps as the chat model
A chat's main model can now run through AI apps already on your PC: OpenCode, Claude Code and Grok Build. Your own plan pays, QuorX holds no key, and QuorX's tools work the same. How to switch them on, and what changes.
If you already pay for an AI app on your computer, QuorX can use it as a chat's main model. The app makes each call with your own plan, subscription or sign-in; QuorX keeps everything else: its files, browser, terminal, artifacts, checks and schedules. It's all in ⚙ Settings → Apps on this PC.
| App on your PC | What it brings | Signed in how |
|---|---|---|
| OpenCode | Every provider you connected in OpenCode | OpenCode's own sign-ins and keys |
| Claude Code | The models of the account it's signed in to | Claude Code's own sign-in |
| Grok Build | Grok's models, for text and build work | Grok's own sign-in (grok login) |
Switch one on
Tick the app in ⚙ Settings → Apps on this PC, then press Test the connection. It lists the models the app offers and costs nothing. QuorX holds no key for these apps: the app makes the call, and your plan pays.
Then pick a model in the chat's model button. The picker now starts with where a model runs: OpenRouter, OpenCode, Claude Code or Grok Build, each with its logo. Pick one, and that source's own filters follow, so a long list stays short.
OpenCode's own free models can't be picked. The free models from OpenCode's own provider answer only inside the OpenCode app, so QuorX shows them in the picker but locked: hover or click one and it says why, and a call to one is refused before it goes out. Free models from other providers you connected in OpenCode work as usual.
No OpenRouter key? You can still chat
A chat can run on one of these models with no OpenRouter key at all. The helper, the reviewer, the prompt enhancer and fresh eyes can run on a model from an app on this PC too, so they keep working without a key. The deciders and paid media stay on OpenRouter: without a key they're off, and the chat says so. Add an OpenRouter key later and they come back.
For the reviewer, pick a model that sees pictures: it looks at screenshots of what was built, so a text-only model such as Grok makes a weak reviewer.
When a call fails
QuorX never swaps your model behind your back. If a call fails (a used-up plan, an expired sign-in, a free tier that ran out), the reply stops and says which model failed and why; the work so far stays. Pick another model and carry on. If you'd rather it carried on by itself, turn on a fallback to OpenRouter in Settings. How OpenCode's failures and fallbacks work, step by step, is in Run a chat through OpenCode on your PC.
When a setting in your provider account blocks a model, QuorX says so straight away and points you to the setting, instead of reporting a server error after several retries. For example, OpenCode Go's "This Go model requires Global regions" points you to Privacy → Global in your OpenCode workspace.
Grok Build: text and build work
Grok Build models read text only, so computer use is off with them, and the chat says so. Everything else (writing, building, running and checking code) works as usual.
Safe by design
These apps can't call QuorX's tools directly, so their models write each call as text and QuorX reads it back. That text could also come from somewhere else (a web page the agent read, or a file), so QuorX only runs a call that:
- carries the key for that step, which changes every step;
- sits on a line of its own;
- is outside code, quotes and plans.
A page or a file can't trigger one. If a model stalls or writes a call in the wrong place, QuorX asks it again, and a reply that loops or runs on is stopped.