Show HN: CSS Studio. Design by hand, code by agent
TL;DR Highlight
A design tool where visually editing CSS directly in the browser allows an AI Agent via MCP to modify the actual codebase, enabling a WYSIWYG workflow regardless of the framework.
Who Should Read
Frontend developers who collaborate with designers in framework-based projects or want to reduce CSS modification work. Also, UI/UX designers who want to directly apply design changes without code.
Core Mechanics
- CSS Studio operates as a design tool inserted as a JS package into the browser. It can be used immediately by adding the package to the development server without a separate SaaS dashboard, and it was confirmed to work smoothly with Vite-based builds.
- When a user visually adjusts the style by clicking on elements in the browser, the changes are transmitted to the AI Agent via the MCP (Model Context Protocol, a protocol for connecting AI with external tools). The AI Agent directly modifies the CSS files in the actual codebase based on this information.
- The core architecture is 'Client-side Injection JS → MCP flow'. Some developers evaluated that the idea of passing visual editing events from the browser to the AI via MCP could be applied to other projects.
- The price is $99, which sparked controversy in the community. Questions were raised about whether this price was appropriate given that the entire Figma suite costs $160, and opinions were raised that a review of the pricing structure was necessary.
- The lack of Tailwind CSS integration was cited as a drawback. In projects using Tailwind, styles are applied as classes, causing only default values to be displayed in the style view. Feedback was received that integrating Tailwind class detection and enabling the Agent to edit based on classes would be a killer feature.
- It was pointed out that there is no diff view showing which code changes before publishing changes. Opinions were expressed that deploying without knowing what code will be modified is unstable in actual production use.
- Concerns were also raised about responsive support. A developer creating a marketing site for an agency asked 'how changes are reflected when dragging and resizing the viewport', but there was no clear answer.
- Several comments criticized the landing page itself as having a generic style that seemed to be generated by an LLM. The opinion was that the team creating the design tool should demonstrate originality directly on the landing page, and there were also bug reports that the layout was broken on mobile.
Evidence
- A developer who actually applied it to a production site using the Motion library said, 'I really like the concept that you can just add a package without a clunky SaaS'. However, they specifically requested improvements such as the absence of a diff view, the unclear purpose of the draw function, and the lack of Tailwind support.
- There was a question, 'If the AI is just applying changes, why use AI instead of directly modifying with a WYSIWYG editor?' This can be understood in the context that using AI as an intermediary is a core value because browser devtools cannot modify the actual source file in framework-based projects.
- Although there were several comments that the $99 price was too expensive, there were also rebuttals. A developer who struggled with Claude for 6 hours and barely managed to create a 'passable' level of similar tool shared their experience, saying that they had newly realized how much invisible detail tools like Figma and Framer handle, and that the price was understandable.
- In the case of using pure HTML/CSS/JS without a framework, the browser's built-in devtools already perform a WYSIWYG role, so the need for this tool is low. On the other hand, if this tool solves the problem of losing the WYSIWYG function of the built-in devtools when using a framework, it would be attractive.
- A developer who operates GPT, Gemini, and Claude models in production asked, 'Which model is best for handling the studio skills of CSS Studio?' This suggests that this tool is not dependent on a specific AI model and can be linked with multiple Agents.
How to Apply
- In framework-based projects such as React, Vue, and Svelte, when a designer wants to directly adjust CSS, adding the CSS Studio package to the development server allows the AI Agent to reflect visually modified content in the actual source file without a developer.
- If you are not using Tailwind, you can consider introducing it now. However, Tailwind projects currently have problems with properly recognizing class-based styles, so it is best to wait for official Tailwind support to be added or submit an issue requesting the feature.
- You can refer to the client-side JS insertion → MCP server connection architecture itself and apply it to create a custom tool that transmits browser UI interactions as context to the AI Agent. There were also developers in the community who said they would apply this pattern to their own projects.
- Considering the lack of a diff view before applying changes to a production site, test it first in a staging environment and use a workflow that compares the code before and after changes with Git to prevent unexpected code changes.
Terminology
Related Papers
Retrying vs Resampling in AI Control
Claude Code처럼 의심 행동을 막고 재시도하는 방식이 오히려 공격자에게 힌트를 줘서 더 위험할 수 있다는 연구.
Mitigating Provenance-Role Collapse in Long-Term Agents via Typed Memory Representation
LLM 에이전트의 장기 메모리가 출처를 뒤섞는 문제를 '타입이 있는 메모리 원자' 구조로 해결한 논문
Push Your Agent: Measuring and Enforcing Quantitative Goal Persistence in Long-Horizon LLM Agents
LLM 에이전트가 '100개 찾아줘'를 실제로 100개 찾을 때까지 멈추지 않게 만드는 방법과 벤치마크.
CoSPlay: Cooperative Self-Play at Test-Time with Self-Generated Code and Unit Test
Ground Truth 없이도 코드와 Unit Test가 서로 평가하며 함께 품질을 높이는 추론 시간 최적화 프레임워크
Multi-Stream LLMs: new paper on parallelizing/separating prompts, thinking, I/O
현재 LLM이 입력·사고·출력을 순차적으로만 처리하는 구조적 한계를 지적하고, 각 역할을 별도의 병렬 스트림으로 분리해 동시에 처리할 수 있는 Multi-Stream 방식을 제안한 논문이다. 에이전트의 효율성·보안·모니터링 가능성을 모두 개선할 수 있다는 점에서 주목받고 있다.
HarnessAPI: A Skill-First Framework for Unified Streaming APIs and MCP Tools
FastAPI HTTP 엔드포인트와 MCP 도구를 하나의 폴더에서 자동으로 동시에 만들어주는 Python 프레임워크