Screenshot helper Runs on our GPU
Where do I click? Find the button on a screenshot and get the exact point to click.
1 credit · under a minute on our GPU
Upload a screenshot and say what you want to do. UI-TARS finds the right control and returns its pixel position, drawn on your screenshot, and can quote the text it sees there.
What you can do
Each example is a real run on our GPU: what went in, the exact instruction, and the result exactly as the model made it. Pick one to see it, then start from the same instruction.
Only the instruction: no file.
What we asked
You are a GUI agent. Output the action to perform the task on this screenshot. Task: click the Save button. Reply in the form: Action: click(point='<point>x y</point>') with x,y in pixels of the screenshot.
What came out, untouched
Action: click(point='(181,105)')
3 s on our RTX 4060 1 credit ($0.10) · One screenshot
Only the instruction: no file.
What we asked
You are a GUI agent. The user wants to convert a PDF to a Word document on this website. Think step by step: which element do you click first, and what are the next two steps? Then give the first action as: Action: click(point='<point>x y</point>').
What came out, untouched
Action: click(point='(879,504)')
2 s on our RTX 4060 1 credit ($0.10) · One screenshot
Only the instruction: no file.
What we asked
This is a phone screenshot of a web tool. 1) What is the page's main action button and where is it (x y)? 2) Is there a privacy or 'files never leave your device' statement visible? Quote it. 3) Output: Action: click(point='<point>x y</point>') for the main action button.
What came out, untouched
1. The main action button is 'Choose a PDF' and it is located at the center of the page, slightly below the middle. 2. Yes, there is a privacy statement visible on the page. It states: 'Files never leave your device' and 'No watermark'. 3. The action button is 'Choose a PDF' and the coordinates are approximately (209,538).
4 s on our RTX 4060 1 credit ($0.10) · One screenshot
3 examples from our own runs, with the times we measured. Open "Prompt and settings we used" under any of them for the exact input.
Action: click(point='(181,105)')
Click Save in a desktop app
The red circle is the model's click: on the Save button.
Prompt
You are a GUI agent. Output the action to perform the task on this screenshot. Task: click the Save button. Reply in the form: Action: click(point='<point>x y</point>') with x,y in pixels of the screenshot.
Action: click(point='(879,504)')
Find PDF to Word on toolboxhere.
Hit the PDF to Word tile. Asked for a 3-step plan, it gave only the first click.
Prompt
You are a GUI agent. The user wants to convert a PDF to a Word document on this website. Think step by step: which element do you click first, and what are the next two steps? Then give the first action as: Action: click(point='<point>x y</point>').
1. The main action button is 'Choose a PDF' and it is located at the center of the page, slightly below the middle. 2. Yes, there is a privacy statement visible on the page. It states: 'Files never leave your device' and 'No watermark'. 3. The action button is 'Choose a PDF' and the coordinates are approximately (209,538).
Phone page: main button and privacy line
Found 'Choose a PDF' and quoted the privacy line correctly.
Prompt
This is a phone screenshot of a web tool. 1) What is the page's main action button and where is it (x y)? 2) Is there a privacy or 'files never leave your device' statement visible? Quote it. 3) Output: Action: click(point='<point>x y</point>') for the main action button.
The prompt and the model decide the output; bigger models and settings cost more credits. These notes come from what worked, and what did not, in our runs.
Times are end to end on our GPU (an RTX 4060 Laptop GPU, 8 GB), model loading included, before any jobs ahead of yours. "Estimate" marks an option we have not measured yet.
| Option | Credits | Typical time | Slowest seen | What to expect |
|---|---|---|---|---|
| One screenshotRecommended | 1 | 49 s | 50 s3 runs | Measured: 2.2-3.6 s per answer plus 46 s model load. |
You answer for your prompt and your choice of model and settings; the safety filter still checks every job, and a refused prompt costs nothing.
1 creditunder a minute on our GPU, plus any jobs ahead
You'll have 7 days to download it; then it's deleted.
Results are kept 7 days. Credits come back only if the job fails or never starts.
Reading the line…
Credits are held when you send a job and spent when the result is ready. They come back automatically if the job fails, is refused or has not started within 24 hours. Results are kept 7 days, or 48 hours after your first download, then deleted; a result that expires unused is not refunded. Pricing and refunds
No. It tells you and shows you where to click on your screenshot.
Yes; the phone trial found the main button and quoted the privacy line correctly.
In our trials the point landed on the right button every time.
At most 7 days, then deleted. Avoid screenshots with private data.
The click point as JSON and your screenshot with the point marked.