A computer task can need several kinds of work: understand the request, choose the next action, find the right control, and write the final result. One model does not have to handle every part in the same way.
OGAD (Off Grid AI Desktop) has a beta Computer Use workflow that can assign those roles to different models. On Mac, you can keep the selected models local and ask the app to carry out a task through the visible computer interface while you supervise it.
[Get OGAD](https://getoffgridai.co/desktop/) | [Read the beta release notes](https://github.com/off-grid-ai/OGAD/releases/tag/v0.0.52-beta.102)
Use 0.0.52-beta.103 for the controls in this guide. The 0.0.52 beta series introduced the combined model-role workflow. The beta also makes task tools available without a Pro entitlement. You still need compatible downloaded models and enough memory for the selected combination.
Start with a job whose result you can inspect. For example, prepare a short comparison document from facts you provide:
Open a new TextEdit document. Compare Option A and Option B using only these notes: [paste notes]. Add a heading, a short comparison, and three questions still unanswered. Save the document as “Options review” in my Documents folder. Do not invent costs or missing details.
The task combines understanding the notes, writing useful text, operating an app, and checking a saved result. Watch the first run so you can catch a wrong action before it carries into the document.
Keep the first input small enough to inspect. Once that task works with your model combination, expand the job rather than beginning with hours of unattended computer work.
The beta's Computer Use settings offer several strategies. The most complete combination is Decision + Reasoning + Specialist.
| Role | Its part in the task |
|---|---|
| Chat model | Writes text and helps resolve uncertain choices. |
| Decision model | Chooses bounded actions from compact Accessibility state. |
| Grounding specialist | Helps with visual recovery and locating screen targets. |
The decision model does not receive the same full screenshot workflow as the visual specialist. These roles are coordinated parts of a task, not several independent agents that you must message separately.
A simpler Reasoning + Specialist strategy combines the Chat model with a specialist in the vision workflow. It can be a useful starting point if you do not have a separate decision model ready.
Multiple models are not automatically faster or more accurate. They add their own memory and compatibility needs. Pick a combination for the job and inspect its result.
Use a downloaded local Chat model. In the Computer Use model catalog, install a compatible specialist and, for the decision strategy, a Decider model offered by the catalog. The role selectors list compatible installed choices. If a selector has no usable local model, use Open Computer Use catalog and complete the relevant download first. A normal chat model is not automatically a replacement for every specialist role.
On Mac, review Settings → Setup & health → System permissions for Accessibility and Screen Recording. Computer Use needs the relevant permissions to inspect and operate the visible interface.
Keep all selected role models local for local inference. A remote role uses its saved server connection and changes where that part of the task is processed. The app and model downloads require setup before offline use; websites involved in a task still need internet.
Computer Use and Web Use have separate strategy settings. In this beta, selecting a Decider for Web Use does not make its browser loop use that model in the same way as Computer Use. Follow the Computer Use route for this guide.
The task view reports progress and retains available activity for review. The optional Picture in Picture setting keeps task controls visible above the app being controlled. It is off by default, so enable it if that suits your work.
If the task stops or repeats a wrong action, stop and inspect the current app state before continuing. A saved progress stage helps preserve context, but it does not make an unfinished document complete. Check the output rather than relying only on the reported status. Try the OGAD beta with one small, supervised local document task. Choose a supported model combination, define the result, and check the saved file. That gives you a useful basis for deciding which longer task to hand over next.