“Local AI” is useful only when you know what stays local.
With a downloaded local model, Off Grid AI can process supported tasks on your device. OGAD (Off Grid AI Desktop) and OGAM (Off Grid AI Mobile) also offer optional connections and device workflows, which have different data paths. Check the model, tool and destination together. A local model does not make a web request or remote service local.
| Get Off Grid AI | Read the privacy policy |
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@offgridmobileai.co, join our Slack community, or talk to us on Reddit.
What happens in a simple local chat?
You select a downloaded model and provide text. The model runs using the device’s resources and produces a reply. The app keeps the conversation in its local data store. This is the straightforward route to use when you want help with a draft without sending the draft to a hosted model.
You still decide what to provide and what to keep. A local answer can contain mistakes, and local storage can contain sensitive material. Processing on your own computer changes the destination of the task; it does not remove the need to review answers or protect the computer.
Start with a non-sensitive example so that you can understand the workflow before adding private work.
Which choices can involve another system?
Look at the action the app is taking, rather than only the word “local” beside the model.
| Your action | Relevant destination |
|---|---|
| Use a downloaded local model | The device running that model |
| Download a model | The model’s download host |
| Search the web | The search provider receives the query |
| Read a live URL | The website receives a request |
| Use a remote model | The configured model host |
| Connect an external service | That service’s supported interface |
| Sync selected data | The paired device receiving it |
| Send a support message | The support or community service you use |
Some of these operations are useful. The point is to choose them deliberately. For example, checking a public product specification can require the web, while rewriting your private notes can stay in a local chat.
The desktop code implements web retrieval separately from local model inference. Its search tools make external requests, and its URL reader fetches the requested page. Tool implementation.
What if you use a model on your home computer?
The request goes to that computer. If you own and manage the host, that gives you a different arrangement from using a third-party hosted model, but the request still leaves the client device.
Check the configured endpoint and how the host is reached. A laptop using a desktop model on the same network is a local-network workflow. Reaching that host while away requires a usable remote network route. Private IP support and Tailscale do not mean that no internet transport is involved.
This distinction matters when explaining the setup to a colleague: “processed on our office computer” is different from “processed on this phone.” Use the accurate description for the selected route.
What stays in captured work history?
On supported Pro desktop workflows, enabling background capture can create retained screen samples and derived work context. Meetings, Clipboard and imported documents have their own input and storage paths. Do not treat them as one switch that silently collects everything.
Capture is opt-in with a visible state. Keep it enabled only for the work you intend to retain, use the available exclusions and check a harmless sample. A pause stops later capture; it does not erase records that already exist.
Choose local processing resources when you want the analysis to remain local. If you configure an external model or connector for a task, review that route separately. The safest description of a setup is the one you can verify from its actual selections.
What changes when devices sync?
Sync creates another copy or shared record on a device you paired and allowed to receive that data. The destination device therefore becomes part of your data setup.
Review the supported data categories and transfer controls. A work laptop and a personal phone may need different choices. You do not have to treat every available sync option as appropriate for every device.
Also distinguish live sync from backups. An operating-system backup, exported file or folder copied elsewhere can create additional copies outside the app’s immediate storage. Removing one local record should not be assumed to remove every independent copy.
Does Off Grid receive any information at all?
The privacy policy describes ordinary captured content as local to your devices. It also identifies separate situations such as payment and license processing, communications, optional diagnostics and basic analytics. Optional third-party services operate under their own policies. Off Grid AI privacy policy.
For that reason, “all use of the app produces zero network traffic” is too broad. It would also hide the choices you can control. Distinguish the content of a local AI task from installing software, purchasing a license or sending a support request.
If you send logs or screenshots to support, inspect what you attach. Describe the issue with a small sample where possible. The same care applies to Slack and Reddit: those are communication services, not extensions of your private local database.
How can you check your intended workflow?
Use a short acceptance test that matches the work you plan to do.
- Select a local downloaded model.
- Leave remote providers and online tools out of this test.
- Paste a harmless note and ask for a summary.
- Check that the answer preserves the note’s facts.
- Disconnect and repeat another short local request.
If it works, you have checked that particular task without a network. You have not tested every app feature, all future settings or every background service on the device.
For a connected workflow, write down the extra destination. For example: “The public search query goes to a search provider; the saved client note stays in local chat.” That record makes the boundary understandable to someone else using the setup.
How do you manage retained desktop data?
OGAD’s data controls distinguish categories such as chats, memory, captures and meetings. Review the scope of the chosen deletion or retention action before applying it. Installed model files are managed separately from personal data in the desktop deletion flow.
Keep useful source material long enough to check an AI answer. If you remove the source, a retained summary is not a replacement for the original evidence. At the same time, avoid retaining sensitive work without a reason.
Download Off Grid AI and start with one local task you can verify. Add connections only when you know what they do and where the relevant data goes.