You’ve been there. You type out a perfect prompt explaining that the button is three pixels off, that the layout breaks at 768px, that the hover state is invisible. Your AI assistant nods politely and generates code that still doesn’t fix it. Why? Because it’s blind.
We’ve spent years obsessing over making AI write better code. Larger context windows. Smarter reasoning. More tokens. But the dirty secret of every AI coding tool — Claude Code, Copilot, Cursor — is that they operate in a sensory vacuum. They can read your codebase, but they cannot see the page they’re building.
That changes today.
Peek-CLI is an open-source tool that lets Claude Code see the browser — literally. It captures screenshots, inspects live DOM state, and feeds the visual runtime back into the AI’s context. No more “imagine the layout” or “guess why this margin collapsed.” The AI sees what you see.
The biggest bottleneck in AI-assisted development isn’t code generation — it’s perception.
I’ve spent the last month debugging a CSS grid issue that only appeared at 800px viewport width. Every time I described it to Claude, it would suggest a fix that worked in theory but broke something else in practice. Turns out, the problem wasn’t my code — it was that the AI couldn’t see the floating sidebar overlapping the main content because the tooltip triggered a resize event. A human would catch that instantly. Claude, blind, needed seven rounds of back-and-forth.
With Peek-CLI, you hand the AI a real-time visual state. You say, “Look at this page now.” It sees the overlapping sidebar. It sees the weird 2px gap on hover. It sees the responsive collapse at 480px. The debugging loop shrinks from minutes to seconds.
This isn’t a theoretical future. This is a CLI command that works today.
Here’s how simple it is: you run a background service that connects to your browser via CDP (Chrome DevTools Protocol). Peek-CLI strips out the noise — only the rendered state, the computed styles, the accessibility tree, and a screenshot if you want. It packages that into a structured observation that Claude can consume. The AI doesn’t need a big vision model; it just needs the right data.
Yes, there are tradeoffs. You lose some abstraction — Peek-CLI is opinionated about what the AI should “see.” But that’s the point. We’ve been drowning in abstraction, pretending that larger context windows would make up for the fact that AI can’t look at a live page. Neutrality is death. Peek-CLI picks a side: the AI should see your browser, not just your code.
If you’ve ever spent thirty minutes explaining a layout bug to an AI that can’t see it, you know the frustration. Peek-CLI gives you back control. It turns the browser into a first-class sensor for your AI assistant. No more blind coding.
Try it. Run `peek` in your terminal. Watch Claude finally see what you’ve been looking at all along.
FAQ
Q: Does this work with any AI coding assistant, or only Claude Code?
A: Right now, Peek-CLI is tailored for Claude Code because of its structured observation format. But the underlying browser sensor layer is tool-agnostic — you could adapt it for Copilot or Cursor with some engineering. The core idea of feeding live DOM state into an AI works regardless of the assistant.
Q: What's the practical implication for my daily workflow?
A: You stop writing 'imagine the page looks like X' prompts. Instead, you capture the actual rendered state and inject it into the AI's context. Layout bugs, responsive breakpoints, hover animations — the AI sees them in real time. Expect to cut debugging time by at least half for visual issues.
Q: Isn't this just a crutch? Shouldn't AI learn to reason about code without relying on visual snapshots?
A: That's the ideal endgame, but today's AI models are fundamentally text-based. They reason through code, not pixels. Telling them to 'imagine the page' is like asking a novelist to describe a photograph they've never seen. Peek-CLI bridges a real gap now, while we wait for true multi-modal models that can render and self-correct. Pragmatism beats purity.