onUI: annotate live web UI and turn it into structured context for AI agents

[iota31]

Most AI-assisted frontend work still starts with a screenshot and a paragraph.

That works, but it is a lossy way to explain UI problems:

• which exact element matters
• how multiple elements relate
• what behavior is wrong
• what should change next

onUI explores a different model:

Website: https://onui.onllm.dev/
GitHub: https://github.com/onllm-dev/onUI

onUI lets you:

• annotate elements directly on the page
• draw regions for spacing/layout issues
• export structured context
• keep everything local with no backend
• optionally connect a local MCP flow so the agent can re-read fresh annotations

The workflow we care about is:
annotate -> agent reads -> fix -> re-annotate -> refine

The interesting question for me is not just whether annotation is useful.

It is:
what is the minimum UI context an agent needs to act reliably?

Would love thoughts on:

1. What data should always be attached to a UI annotation?
2. Should agent-facing output prioritize element targeting, layout relationships, or natural-language intent?
3. Does this feel more like a frontend devtool, QA tooling, or a new category?