Glue Gives Coding Agents Eyes So They Can Actually See What They Are Building

The Macro: Coding Agents Build Blind

Coding agents write code. They do not see the result. When you ask Claude Code or Cursor to change a button’s color, add padding to a layout, or redesign a component, the agent modifies code and hopes the visual output matches your intent. It cannot see the rendered UI to verify.

This creates a tedious feedback loop. The agent makes a change. You refresh the browser. You look at the result. You describe what is wrong. The agent makes another change. You refresh again. Each cycle takes minutes, and getting a visual design right might require dozens of iterations.

The core issue is that coding agents lack a visual layer. They work in text, but UI development is fundamentally visual. Telling an agent “move the heading 20 pixels down and increase the margin between cards” is slower and less precise than just pointing at the element and dragging it.

Figma solved design-to-code handoff for human teams, but it does not work with coding agents. The bridge between visual design intent and code generation is missing.

Glue, backed by Y Combinator, provides the visual layer for coding agents. It is an open source design canvas that integrates with coding agents, giving them the ability to see and manipulate UI visually.

The Micro: Comment on Your UI and Watch the Agent Fix It

Perbhat Kumar (CEO, ex-Amazon and ex-Microsoft) and Tejas Priyadarshi (CTO, ex-Microsoft and ex-Meta) built Glue as a bridge between visual design and AI-powered code generation.

The product has two key features. First, in-UI commenting lets developers leave feedback directly on app elements. Click on a button, type “make this larger and change to primary color,” and the connected coding agent converts that visual annotation into actual code changes. Second, the Glue Stick feature lets developers browse any website, select a component they like, and have Glue automatically adapt it to their codebase.

The setup is minimal: a single command (glue init) compatible with any framework. It works with Claude Code, OpenAI Codex, and other coding agents.

The open source approach is a strong strategic choice. Developer tools that are open source build trust, community, and adoption faster than proprietary alternatives. It also means Glue can evolve with the rapidly changing coding agent ecosystem rather than being locked to a single provider.

Competitors include v0 by Vercel (AI-generated UI), Bolt (AI web development), and various Figma-to-code tools. Glue differentiates by being agent-native rather than agent-adjacent. It is designed specifically to be the visual layer that coding agents use, not a standalone design tool with AI features bolted on.

The Verdict

Glue addresses a real gap in the coding agent workflow. Visual feedback loops are the biggest friction point in AI-assisted UI development.

At 30 days: how many developers have installed Glue, and what is the daily active usage?

At 60 days: are coding agents producing better UI code when they have Glue’s visual context compared to text-only prompts?

At 90 days: is Glue becoming a standard part of the coding agent setup for frontend development?

I think Glue has the right insight. Coding agents need visual context to build good UI, and no other tool provides it natively. The open source approach and simple setup lower the barrier to adoption. If the integration with major coding agents is seamless, Glue becomes as essential as a browser preview in the modern development workflow.