SUB&SUB Relay
Stop describing screenshots to your AI. Mark them up.
Start new thread

SUB&SUB Relay - Stop describing screenshots to your AI. Mark them up.

by•
Drop a file — or capture a full-page scrolling screenshot with the Chrome extension — redact and annotate right in the browser, and get back a public URL that auto-expires after a lifetime you pick (5 to 30 min). Includes an AI-agent prompt variant — for agents on a remote server, or any chat that can't take image attachments. Box what you want changed, write a note on each, and the agent gets the URL plus a numbered instruction list tied to those regions.

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt 👋 Markup Relay started with a small annoyance I kept running into while working with coding agents. Sometimes the agent is running somewhere I can’t simply drag a file into: a remote machine over SSH, a terminal, or a chat interface that doesn’t support attachments. I’d be looking directly at a broken UI, but had no clean way to show the agent exactly what was wrong. So I built Markup Relay. Drop in a screenshot, draw a box or arrow around the problem, and write something like: “This overflows at 320px.” You get a shareable URL and a one-click **Agent** button. The model receives both the annotated image and a numbered instruction list matched to the annotations: `Box 1 — region at 12%,30%, size 20%×15%: this overflows at 320px` The numbers are embedded in the image and repeated in the text, so the model can connect each instruction to the correct part of the screen instead of guessing. Tools that live inside a browser can already target DOM elements directly. Markup Relay is for everything that isn’t a DOM: * Native desktop apps * Phone and tablet screens * Screenshots of PDFs and Figma frames * Dashboards behind SSO * Photos of physical devices * Screenshots sent by teammates The idea is to provide a markup vocabulary, not another selector: boxes, ellipses, arrows, freehand drawing, highlights, blur, mosaic, numbered callouts, and cropping. “Follow this arrow, then look at 3” is how people have explained visual problems to each other forever, and vision models are getting increasingly good at understanding the same language. Privacy was also important to me. By default, everything outside the areas you mark is blurred. The same gesture that says “look here” also says “ignore everything else.” All editing happens locally in the browser, and the original unedited image is never uploaded. Because it’s just a web page, there’s nothing to install. You can annotate something on your phone and paste the link into an agent running on your desktop. It also handles video. Oversized clips are converted locally into the frames where the picture changes, plus a contact sheet — the original video never uploads. Uploads are served as downloads rather than executable web content — with one deliberate exception: files verified to be real images are served inline, so a model can actually look at them. SVG is never treated as an image. Links expire automatically. It’s free with no sign-up: up to 3 MB per file, one file at a time. Plus supports 10 MB files, batches of up to 20, and longer link lifetimes. I’d love feedback, especially on the annotation-to-instruction-list workflow. That’s the part I spent the most time getting right.