Problem: How do I quickly tell AI exactly which elements I want to change? Solution: Agentation This npm package displays a toolbar in your app so you can click on elements, add comments, and then send those comments to your LLM (like Claude Code).

We were inspired by the agentation UX and added a similar experience in Frontman, which allows you a visual agentic experience integrated to your Astro/Vite/WordPress/NextJS project

There's also an MCP server that provides more streamlined integrations and a watch mode. Quite slick. More:

Still need to try honestly. Been playing with Codex lately and the internal browser with annotation has been enough. Even thoug this one you're suggesting will work cross apps, and I could use it in OpenCode 😋 Thanks for sharing

Cool to see an open (OK, semi-open) solution. Codex has this built into the native browser and it works incredibly well.

Tidewave does this and a lot more. Check it out.

I’m curious, if you verbally mention an element doesn’t the agent know what you are referring to usually? I’ve had success with this approach so far.

Yes, that's generally my approach and I've found it works pretty well. Agentation is useful in two cases: 1. I want to put in a lot of visual feedback quickly and not waste time trying to describe each target element. 2. The element is hard to target uniquely via words

This is exactly the same as react-grab

In Cursor 3, there is a design mode where we can do exactly this! I love using that to pinpoint UI changes

I tend to just grab a screenshot and paste it in the terminal. This works pretty well, is fast, and uses essentially native functionality that doesn't require me running a bunch of extra software.

This is such a smart UX pattern. Being able to point-and-click on UI elements and send structured comments to an LLM feels like the missing bridge between ‘chat about my app’ and ‘actually change my app.

This is really helpful, thanks man

Pointing at the thing beats describing it in prose every time. AI coding flows have been missing this layer. how does it handle state that changes mid-session?

How is this better from Cursors "design mode", Visual manipulation in built in browser?

it’s great

Neat approach, saves time tagging UI elements visually

very well done.