html-anything is a local-first HTML editor that integrates multiple AI coding agents directly on your machine, routing composable skill templates through whichever agent CLI you have installed. It’s a technical approach to agent-driven content creation that avoids cloud API calls by leveraging locally available AI CLI tools.
How html-anything orchestrates AI coding agents for HTML generation
At its core, html-anything is an agentic HTML editor combining 75 composable SKILL.md templates with eight distinct CLI coding agents it auto-detects on your PATH. These agents include Claude Code, Cursor Agent, Codex, Gemini CLI, Copilot CLI, OpenCode, Qwen Coder, and Aider. Instead of relying on a single AI backend, it dynamically routes skill prompts to whichever agents are available locally.
The editor supports generating content for nine different deliverable surfaces: magazine articles, keynote decks, résumés, posters, Xiaohongshu cards, tweet cards, web prototypes, data reports, and Hyperframes videos. Each surface has multiple skill presets, enabling tailored content generation workflows.
Under the hood, html-anything streams generation results via Server-Sent Events (SSE) into a sandboxed iframe, providing a live, real-time preview that feels like watching someone type out the content. The final output can then be exported with one click, with CSS inlined for platforms like WeChat, X (formerly Twitter), Zhihu, or downloaded directly as HTML or PNG.
The whole architecture is designed to be zero-API-key and local-first. It leverages existing CLI sessions for the agents, so there’s no additional cost or cloud dependency. This is a distinctive approach compared to many AI-powered editors that depend on centralized API calls.
Technical strengths and architectural tradeoffs
The standout technical strength is the multi-agent orchestration model via composable SKILL.md templates. By decoupling the skill definitions from the agents themselves, html-anything achieves a flexible, extensible system where the same skill can be routed through different AI engines depending on what’s locally available.
This approach requires the editor to auto-detect the presence of CLI agents on the user’s PATH and manage routing logic accordingly. It’s a non-trivial engineering challenge to maintain consistent prompts and outputs across different agents with varying capabilities and APIs.
The streaming generation via SSE into a sandboxed iframe is another strong point. It improves the developer and user experience by providing immediate, incremental feedback on content generation, rather than waiting for a bulk response. This technique also enforces security boundaries by sandboxing the rendered HTML.
However, this architecture has tradeoffs. Relying on local CLI agents means the user must install and maintain these tools, which can be a barrier for less technical users. The diversity of agents also means handling different response formats and error modes, adding complexity.
The composability of 75 SKILL.md templates provides rich customization but also a learning curve. Users need to understand the skill system to get the most out of it. The zero-API-key model also limits access to cloud-powered agents unless they are installed locally.
Overall, the codebase is surprisingly clean for a project managing so many moving parts. The modular skill templates and agent detection logic are well organized, and the live streaming UI is smooth and intuitive.
Quick start
git clone https://github.com/nexu-io/html-anything
cd html-anything
pnpm install
pnpm -F @html-anything/next dev
These commands clone the repo, install dependencies using pnpm, and start the development server for the Next.js frontend. From there, you can open the editor in your browser and start experimenting with available agents and skill templates.
verdict
html-anything is a sophisticated tool for developers and content creators who want agent-driven HTML generation without relying on cloud APIs. Its local-first model and multi-agent routing architecture make it unique among AI-powered editors.
The zero-API-key approach reduces cost and privacy concerns but requires some setup to install CLI agents. The breadth of deliverable surfaces and composable skill templates offers powerful customization but a steeper learning curve.
If you’re comfortable managing CLI AI tools and want an extensible, streaming HTML editor that integrates multiple agents under one roof, html-anything is worth exploring. For casual users seeking a turnkey AI writing tool, this approach may feel heavyweight.
The project’s architecture and streaming approach offer useful patterns for anyone building local-first AI-assisted content tools or multi-agent orchestrators.
Related Articles
- Mapping the landscape of terminal-native AI coding agents: a curated directory analysis — A curated directory catalogs over 80 terminal-native AI coding agents and harnesses, highlighting open-source projects,
- Softaworks Agent Toolkit: A modular plugin marketplace for AI coding agents — Softaworks Agent Toolkit offers 40+ modular AI skills and plugins for coding agents like Claude Code, enabling composabl
- SkillUI: Automating design system extraction into Claude Code skills — SkillUI reverse-engineers design systems from URLs or repos into Claude Code skill packages using static and visual anal
- ok-skills: a universal skill pack for AI coding agents — ok-skills provides 38 reusable SKILL.md playbooks for AI coding agents like Claude Code and Cursor, standardizing cross-
- 10x CLI coding agent: tiered AI model routing for faster coding workflows — 10x is a TypeScript CLI coding agent that speeds coding up to 20x by routing tasks across a tiered AI model system with
→ GitHub Repo: nexu-io/html-anything ⭐ 4,494 · HTML