Design Atlas
Type
component registry
Verdict
Niche
Licence
Ship licence, open source permissiveopen-source-permissivefree; MIT (repo shadcn-labs/agentcn, about 480 GitHub stars at review). Recipes call paid third-party APIs you bring keys for.
Agent channels
llms-txt, registry, api, skill
Pricing
free
Reviewed

← Atlas · Topics: agents-and-prompts, ai-interfaces, design-md

agentcn ​

Type
component registry
Verdict
Niche
Licence
Ship licence, open source permissiveopen-source-permissivefree; MIT (repo shadcn-labs/agentcn, about 480 GitHub stars at review). Recipes call paid third-party APIs you bring keys for.
Agent channels
llms-txt, registry, api, skill
Pricing
free
Reviewed

What it is ​

agentcn applies the shadcn copy-the-source model to backend AI agents. It is by Aniket Pawar under the community Shadcn Labs organisation, the same maker as mcpcn. The docs say plainly that it is not a UI component registry: each recipe installs an agent definition, an instructions.md, tool files and sometimes a SKILL.md into your project. There are 19 recipes, each written for four frameworks (Vercel's Eve, Flue, Mastra, and LangGraph via the Dawn meta-framework), for 76 registry items. It launched in June 2026 and added Mastra and LangGraph in July.

When to open it ​

When the interface is done and you need a working agent behind it: deep web research with citations, chat over a PDF, a YouTube video or a database, a PR reviewer, a Slack bot, or a browser agent. It pairs naturally with an AI chat UI kit, since this covers the logic side and not the screens.

Most useful ​

  • Research: Deep Search (searches, critiques its own findings, repeats until each claim is cited) and Docs Expert.
  • Retrieval: Chat with PDF (page-cited), Company Knowledge (with PII redaction), Docs Chatbot and Chat with Database (read-only SQL).
  • Automation: GitHub PR Review, Slack Agent, Google Sheets, a Playwright Browser Agent and a sandboxed shell "Claw" assistant.
  • Design-adjacent: Extract DESIGN.md turns a domain into a DESIGN.md plus a Tailwind v4 @theme block and CSS variables, following the designmd.supply pipeline with one vision-model call; AI SEO Audit scores about 30 checks and returns a fix prompt.
  • Live previews: each recipe page runs the agent in the browser and shows its tool calls as it works.

Using it with agents ​

Install with npx shadcn@latest add @agentcn/<framework>/<recipe> (for example @agentcn/eve/deep-search), or register the @agentcn namespace once. The shadcn MCP server can browse and install recipes from your editor. Like mcpcn, it serves llms.txt, llms-full.txt, Markdown versions of every page, an OpenAPI file and a site skill at /.well-known/agent-skills/site-skill.md.

Watch out for ​

  • Most recipes depend on outside services: Bright Data for search, context.dev for the DESIGN.md and SEO recipes, and Anthropic models by default. Each needs its own API key and billing.
  • Default model IDs are hard-coded in the agent files; change them to match your provider.
  • It is a young project with one maintainer, and the framework APIs it targets (especially Eve and Flue) are themselves new.
  • Despite the shadcn naming and tooling, it ships no interface components; you still need a UI kit for the front end.

Reusable ideas ​

  • Let a research agent grade its own draft and search again to fill gaps, with a hard cap on rounds.
  • Keep instructions, tools and skills as separate plain files so each can be read and edited on its own.
  • Derive tokens deterministically from measured styles and let the model write only the prose.
  • Show a live trace of tool calls next to any agent demo so users see how the answer was built.

mcpcn, Prompt Kit, designmd.supply, DesignMD.cc, shadcn/ui