An MCP server that checks whether an interface is actually correct for Japan. It reads real CSS and markup, typography, line-breaking, colour, forms, scores them against the W3C JLReq and JIS X 4051 standards, and returns the fixes an agent should apply.
I have been building sites with AI, and the same thing kept happening: the output looks right in English and quietly breaks in Japanese. Line-height too tight for kanji. No Japanese font in the stack, so characters fall back to whatever the OS happens to have. Text set to justify, which browsers cannot do well for CJK. Sub-14px kanji. And on every form, the reversed name order and the single combined phone field. None of it is the model's fault. The conventions just are not in its training data densely enough to beat the English defaults.
Years on Japanese products had given me that checklist in my head. The first version of this tool wrote it down as generators an agent could call: ask for a Japanese form, get correct markup back.
Generating a correct form solved the easy half. The harder half runs the other way. Once an agent has written a page, nothing tells it the page is wrong for Japan. It cannot see its own 1.2 line-height or its justified paragraphs.
So the tool became a checker. Point it at real CSS and markup and it scores them against the actual standards, the W3C Requirements for Japanese Text Layout and JIS X 4051, and hands back every violation with its fix. The form generator is still inside. It is now one tool among the checks.
Some of these rules are law. Line-breaking is specified in the W3C JLReq and JIS X 4051. Others are documented practice, like 1.8 line-height for kanji. Every finding says which it is and cites the source, so the tool never dresses a preference up as a standard. For something an agent will apply without arguing, that line is the whole point.
Producing a correct component is useful. Catching an incorrect one is more useful, because that is the failure that actually ships. The core tool reads code the agent already wrote and returns a score with line-level fixes, so correctness can happen at review, not only at generation.
I did not try to encode all of Japanese web design. I targeted the surfaces where AI output reliably breaks: typography, line-breaking, colour, forms, keigo, trust. Each is narrow enough to be prescriptive about.
Every rule, prefecture, name, and postal code ships inside the package. Running it needs only npm. No auth, no quota, no outbound calls.
Open source on GitHub, published on npm, now at v2.0.0. Validated Premium on the lobehub MCP directory. About 440 monthly downloads as of April 2026. Scaffolded with Claude Code; the tool design, the spec research, and every ship decision are mine.
npm · japan-ux-mcp GitHub lobehubMost design systems live in a document a person reads once. Turning mine into tools an agent could call was the first step. What I did not expect was that the more valuable direction was the opposite one: not producing correct Japanese UI, but catching incorrect Japanese UI in code already written. A style guide that can grade its own output.