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 | Read the beta release notes
What would you like to do with Off Grid AI?
Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.
Write to support@getoffgridai.co, join our Slack community, or talk to us on Reddit.
This guide now follows desktop beta 122. Its core task tools are available without a Pro entitlement. The paired browser connection and Pro services have their own access requirements. Prepare compatible downloaded models and enough memory for the selected combination.
Give the task a clear finished result
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.
What each model role contributes
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.
Prepare the local models and permissions
Use a downloaded local Chat model. Open Models > Tasks. Install a compatible grounding specialist and, for a decision strategy, a decision model offered by the catalog.
The role selectors list compatible installed choices. If a selector has no usable local model, return to Models > Tasks 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.
Check the models assigned to each task
The Tasks tab separates grounding specialists from decision models. Its model lineup shows which models are assigned to the task roles. The reasoning model comes from Text. Check that each required role has a usable local model before starting a task.
The released catalog includes Kev 4B, using the Kev 1.0 checkpoint. This is a Computer Use decision package with its own runtime and base files. It is not an ordinary chat GGUF. Let the model interface obtain the package it requires; do not substitute a similarly named text-model file. An existing compatible base can be reused, including after changing the models folder.
The packaged Kev runtime has platform checks for Apple Silicon macOS, Windows, and x64 Linux. Runtime availability does not establish that every task or permission path works on every operating system. This guide's supervised document example remains a Mac workflow. Read the app's platform and download status before selecting a role.
The paired browser panel also shows Desktop's task models and role controls. A change there changes the paired Desktop configuration. It does not download a second set of models into the extension. Recheck Desktop if another window changed a role while the panel was open.
These controls follow the beta 122 model catalog. Start with one working role combination before adding another model.
Choose a strategy and run one task
- Open the chat's Settings, then Tasks.
- In Computer Use, set Model strategy to Reasoning + Specialist, or choose Decision + Reasoning + Specialist when you have the required decision model.
- Select the installed local Grounding specialist and, when the strategy requires it, the local Decision model.
- Enable the vision path for Reasoning + Specialist. With a decision strategy, keep its required Accessibility path and enable vision recovery if you want the specialist to handle visual fallback.
- Return to chat, enable Assistant, and submit the small document task.
- Watch the task's progress. At the end, open the saved document and check its text and location.
Computer Use and Web Use keep separate strategy settings. Check both after using a decision-model card: its Use action can assign that decision model to both roles. A role selection does not establish that both task engines use it in the same way. Use the Computer Use route for the document example, then test Web Use separately for browser work.
Supervise the task without losing its place
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.
Let each model handle a defined part
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.
Top comments (0)