The tool ships no model
HTML Anything makes a bet most AI design tools avoid: it ships no model and asks for no API key. On startup it scans your PATH for a coding-agent CLI you have already logged in, then hands that agent the job of writing HTML. Claude Code, Cursor Agent, Codex, Gemini CLI, GitHub Copilot CLI, OpenCode, Qwen Coder, and Aider are all detected automatically, including from the odd install directories a GUI-launched Node process usually misses, like ~/.local/bin and /opt/homebrew/bin. You swap between them from a top-bar picker. Because the work runs through the session you already pay for with claude login or cursor login, the project puts the marginal cost at zero dollars.
That framing inverts the usual pitch. Instead of renting you inference, nexu-io/html-anything reuses inference you already have, and keeps the whole loop local.
HTML as the finished form, not Markdown
The reason to bother is an argument the README borrows from Anthropic's Claude Code team, who said they stopped writing internal docs in Markdown and now ship HTML. HTML Anything runs with that idea. Markdown is good for the writer; HTML is good for the reader. Markdown's layout is whatever the renderer decides, while HTML's layout is yours. A Markdown file screenshotted into a tweet looks like a text file, while an HTML page already looks like a designed image. The project's word for the goal is "ship-ready": when generation finishes, the artifact is what your audience sees, with no touch-up pass promised for later.
You feed it Markdown, CSV, Excel, JSON, SQL, or raw notes, press ⌘+Enter, and the local agent returns a single-file HTML page. From there, one click sends it to WeChat, X, or Zhihu, or downloads .html or .png.
75 skills across 9 surfaces
What keeps the output from looking generic is a library of 75 composable skill templates, each defined by a SKILL.md file, spread across nine deliverable surfaces: magazine articles, keynote decks, résumés, posters, Xiaohongshu cards, tweet cards, web prototypes, data reports, and Hyperframes videos. The featured skills are specific rather than vague. deck-swiss-international builds a 16-column grid with one saturated accent across 22 locked layouts. deck-guizang-editorial offers 10 locked layouts across five palettes and reads like a printed art-zine. doc-kami-parchment swaps plain white for a #f5f4ed ground and a single serif voice for long reading. video-hyperframes emits 6 to 10 sequential 1920×1080 frames with hidden duration and transition markers, ready to hand to Remotion for an .mp4.
The constraints inside each SKILL.md are where the project spends its discipline: a CJK-first font stack, an 8 px baseline grid, a minimum contrast ratio, and a rule that says use real data, not placeholders. That last set comes from alchaincyf/huashu-md-html, one of four open-source projects the README credits by name. The others are nexu-io/open-design for the agent-detection layer and the SKILL.md protocol, mdnice/markdown-nice for CSS that pastes cleanly into WeChat and Zhihu, and gcui-art/markdown-to-image for the iframe-to-PNG export path.
Who it fits
HTML Anything is narrow on purpose. It suits people who publish to Chinese platforms like WeChat and Xiaohongshu as much as to X, who already work inside a coding-agent CLI, and who would rather generate a finished page than open a design tool. If you do not already have one of those CLIs logged in, the whole zero-key premise falls apart, and the README is upfront that the sandboxed preview and the export chain are the parts doing the polishing. For the audience it targets, the pitch holds together: the format readers want is HTML, and the agent you already pay for can write it.
GitHub: https://github.com/nexu-io/html-anything
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)