A chart can say more than a page's extracted text. OGAD (Off Grid AI Desktop) can use a local vision model to answer questions about a screenshot attached from its Chrome extension. This helps when the evidence is a graph, a layout, or a visible error message.
The active desktop model must support image input. A text-only model cannot inspect the screenshot. Select a local vision model to process the image on your computer, and keep the source page open for comparison.
Download OGAD | Desktop beta 122 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.
Choose a visual question
Start with a question about what is visible. For a chart, that could be the largest change between adjacent months, the unit on an axis, or a difference between two series. For an error dialog, it could be the exact message and the next step suggested on screen.
Avoid asking the image to prove information it does not contain. A revenue chart cannot establish why revenue changed. A screenshot of a form cannot show what the server will do after submission. The model can describe evidence and suggest checks; it should keep those two activities separate.
Prepare a model that accepts screenshots
Load a compatible local vision model in OGAD. Some model packages need a matching image projector as well as the main model file. Let the app complete the supported package download, then confirm that image input is available.
Install the early-access browser extension and pair it through Settings > Desktop. Compare and approve the six words on the desktop. This guide uses the paired desktop route; prepare the required desktop access and Pro tools before starting.
The extension checks whether the active model supports image input. If the screenshot control is unavailable, changing the wording of your question will not add vision support. Return to the model selection first.
Attach the relevant view
- Open the page and scroll to the chart or error you want to discuss.
- Make the labels readable. Close overlays that cover the evidence.
- Open the extension and use ordinary chat, with Agent and image-generation mode off.
- Choose Attach screenshot from the composer controls.
- Confirm that an attachment appears, then enter a specific question.
If you switch to a model that cannot accept images, the extension can remove the staged screenshot. Check the attachment again after a model change. Sending the question without it produces a different task.
Read a chart without inventing precision
Use a prompt such as:
Describe this chart. State the title, axis units, and series names first. Then identify the largest visible increase. Quote labelled values where available. Mark estimates as estimates, and do not infer a cause for the change.
Compare the first two facts with the image before reading the interpretation. If the vertical axis is thousands of units, a label of 20 means 20,000 units. If the axis starts above zero, the visual height difference can exaggerate the apparent change.
For a useful follow-up, ask for the calculation separately:
Using only the two labelled values, show the absolute change and the percentage change. State the earlier value used as the denominator.
Check the arithmetic yourself or with a calculator. Image reading and calculation are separate opportunities for error. When the chart has no exact labels, get the underlying table before using precise figures in a report.
Use screenshots for a work-app error
A screenshot can also help you understand an unfamiliar dialog. Ask the model to transcribe the visible message and list the checks suggested by that message. Include what you were trying to do in your own words.
For example: “I was exporting a draft. Read this error and distinguish a missing permission from a missing file.” This gives the model the task context without asking it to guess everything from a small image.
Do not put secrets in the screenshot unnecessarily. Close unrelated panels before capturing. A local model avoids a cloud AI endpoint, but the image can still become part of the saved conversation.
Fix the common failures
| Result | Check |
|---|---|
| No screenshot control | The active model and required projector support vision |
| Model says no image was supplied | Attachment still present after mode or model changes |
| Small text is wrong | Enlarge the relevant page area and capture again |
| Answer invents numbers | Ask for labelled values only and use the source table |
| Capture is refused | Page access and whether the browser permits that page to be captured |
The website may need internet, but local image analysis does not require a cloud model once the app and model are installed. Remote model choices change that processing path.
Try OGAD with one readable chart. Verify its labels and units before asking for the interpretation you need.
Top comments (0)