Agent Skills

agent-elements

Use whenever the user wants to build or modify a chat, agent, or tool-calling UI in a React 19 + Tailwind v4 project — especially if the code imports from `@/components/agent-elements/*` or the project has that folder on disk. Triggers: "agent chat", "tool call UI", "streaming chat", "plan approval", "AgentChat", "InputBar", "tool renderer", mentions of Agent Elements, or requests to add a new agent surface with shadcn. Do NOT use for plain chat UIs that don't need tool/plan/approval cards, or f

Install

npx skills add https://github.com/21st-dev/agent-elements --skill agent-elements
SKILL.md

Agent Elements skill

Project-aware context for building chat and agent UIs with Agent Elements — an open-source shadcn registry at https://agent-elements.21st.dev.

What this skill gives you

When this skill loads, you know:

  1. The registry is shadcn-compatible. Every component is installed with npx shadcn@latest add https://agent-elements.21st.dev/r/<component>.json. Files land under components/agent-elements/ (the library's internal components/ prefix is stripped — see Paths below).
  2. The API is typed around the Vercel AI SDK. Messages are UIMessage[] from ai, status is ChatStatus. useChat() plugs in directly.
  3. The full component catalog with API shapes and composition rules (see sections below).
  4. Theming guardrails — the Tailwind tokens Agent Elements depends on.

Detection

Consider this project "Agent Elements-ready" if any of these are true:

  • components/agent-elements/ exists on disk
  • components.json includes an alias or registry reference to Agent Elements
  • package.json dependencies include ai + @tabler/icons-react and the user mentions Agent Elements

If the folder does not exist yet, install on demand with:

npx shadcn@latest add https://agent-elements.21st.dev/r/agent-chat.json

agent-chat transitively pulls every other component it needs via registryDependencies (MessageList, InputBar, tool renderers, shared utils).

Paths (post-install layout)

After shadcn add, files sit under @/components/agent-elements/ with this shape:

components/agent-elements/
  agent-chat.tsx
  message-list.tsx
  input-bar.tsx
  markdown.tsx
  user-message.tsx
  error-message.tsx
  text-shimmer.tsx
  spiral-loader.tsx
  input/
    attachment-button.tsx
    send-button.tsx
    file-attachment.tsx
    suggestions.tsx
    model-picker.tsx
    mode-selector.tsx
  tools/
    bash-tool.tsx
    edit-tool.tsx
    search-tool.tsx
    todo-tool.tsx
    plan-tool.tsx
    tool-group.tsx
    subagent-tool.tsx
    mcp-tool.tsx
    thinking-tool.tsx
    generic-tool.tsx
  question/
    question-tool.tsx
  hooks/use-tool-complete.ts
  utils/cn.ts
  types.ts

Import rule: always import from the exact file, never from a barrel.

// ✅
import { AgentChat } from "@/components/agent-elements/agent-chat";
import { BashTool } from "@/components/agent-elements/tools/bash-tool";

// ❌ — no barrel exists
import { AgentChat } from "@/components/agent-elements";

Component catalog

Chat surface

  • AgentChat — the full chat shell. Renders MessageList + InputBar, handles tool invocations via toolRenderers, shows an empty state with optional suggestions. Props: messages, status, onSend, onStop, toolRenderers?, suggestions?, attachments?, classNames?, slots?.
  • MessageList — transcript only. Use when you need the input bar somewhere else. Accepts toolRenderers and showCopyToolbar.
  • UserMessage / ErrorMessage / Markdown — low-level message pieces. Markdown streams safely (external links get rel="noreferrer" by default).

Input

  • InputBar — composer. Props: status, onSend({ content }), onStop, value? + onChange? (controlled), attachedImages/attachedFiles with their remove handlers, leftActions/rightActions slots, suggestions?, questionBar?, infoBar?.
  • Suggestions — quick-prompt chips for the empty state or inline.
  • ModelPicker / ModeSelector — designed to drop into leftActions. Both accept a simple { id, name, version? } / { id, label, icon?, description? } shape. Do not import CLAUDE_MODELS — it was removed; supply your own array.
  • SendButton / AttachmentButton / FileAttachment — usable standalone if you're building a custom composer.

Tool cards

All tool cards accept a part prop of type Extract<UIMessage["parts"][number], { type: \tool-` }>from the AI SDK. Register them viatoolRenderersonAgentChat/MessageList`:

<AgentChat
  toolRenderers={{
    Bash: BashTool,
    Edit: EditTool,
    Write: EditTool,      // Write reuses EditTool
    Search: SearchTool,
    WebSearch: SearchTool,
    TodoWrite: TodoTool,
    PlanWrite: PlanTool,
    Task: SubagentTool,
    Thinking: ThinkingTool,
  }}
/>

Cards available:

  • BashTool — command + stdout, collapsible.
  • EditTool — diff card. Supports input.old_string/input.new_string or output.structuredPatch, plus an approval footer via input.approval.
  • SearchTool — grouped search results. Pass results or use output.results.
  • TodoTool — diffed todo list from input.todos vs output.oldTodos.
  • PlanTool — plan title + summary with approve/reject footer.
  • ToolGroup — collapses consecutive tool calls into one row.
  • SubagentTool — sub-agent task with nested tools.
  • McpTool — generic MCP tool output; use parseMcpToolType from @/components/agent-elements/tools/tool-registry to get mcpInfo.
  • ThinkingTool — collapsible reasoning row.
  • GenericTool — fallback for unknown tools.
  • QuestionTool — clarifying question with single/multi/text answer kinds.

Streaming states

  • TextShimmer — shimmering status label.
  • SpiralLoader — Lottie spiral; use for multi-second loading states.

Composition patterns

Full chat with tool rendering (most common)

"use client";

import { AgentChat } from "@/components/agent-elements/agent-chat";
import { BashTool } from "@/components/agent-elements/tools/bash-tool";
import { EditTool } from "@/components/agent-elements/tools/edit-tool";
import { SearchTool } from "@/components/agent-elements/tools/search-tool";
import { useChat } from "@ai-sdk/react";

export default function Chat() {
  const { messages, status, sendMessage, stop } = useChat();
  return (
    <AgentChat
      messages={messages}
      status={status}
      onSend={({ content }) => sendMessage({ text: content })}
      onStop={stop}
      toolRenderers={{
        Bash: BashTool,
        Edit: EditTool,
        Write: EditTool,
        Search: SearchTool,
      }}
    />
  );
}

Composer with mode + model pickers

import { InputBar } from "@/components/agent-elements/input-bar";
import { ModeSelector } from "@/components/agent-elements/input/mode-selector";
import { ModelPicker } from "@/components/agent-elements/input/model-picker";
import { IconBulb, IconCursor } from "@tabler/icons-react";

const modes = [
  { id: "agent", label: "Agent", icon: IconCursor },
  { id: "plan", label: "Plan", icon: IconBulb },
];
const models = [
  { id: "sonnet", name: "Sonnet", version: "4.6" },
  { id: "opus", name: "Opus", version: "4.7" },
];

<InputBar
  status="ready"
  onSend={handleSend}
  onStop={handleStop}
  leftActions={
    <>
      <ModeSelector modes={modes} defaultValue="agent" />
      <ModelPicker models={models} defaultValue="sonnet" />
    </>
  }
/>

Custom tool renderer

toolRenderers values are React components that receive { part, chatStatus }. Return whatever UI you want; reuse GenericTool as a fallback shell.

Theming

Agent Elements reads these Tailwind CSS vars (shadcn-style). Do not remove or rename them in the consumer theme:

  • --an-foreground, --an-background, --an-primary-color
  • Standard shadcn tokens: --background, --foreground, --border, --muted, --muted-foreground, --accent, --primary, etc.

Customising a component is just editing the installed file. Prefer that over wrapping — the code is yours now.

When NOT to use Agent Elements

  • Projects using assistant-ui, ai-elements, copilotkit, or another kit — don't mix.
  • Pure chat UIs that never render tool calls or plans — InputBar + your own message rendering may be enough; skip AgentChat.
  • React < 19 or Tailwind < v4 — the components depend on both.

Quick answers for common asks

  • "Add Agent Elements to this project" → run npx shadcn@latest init if components.json is missing, then npx shadcn@latest add https://agent-elements.21st.dev/r/agent-chat.json.
  • "Switch the default SendButton look" → edit components/agent-elements/input/send-button.tsx directly. Tokens live on --an-* CSS vars.
  • "Render a custom tool" → map its type in toolRenderers; fall back to GenericTool for unknown tools.
  • "Use with useChat" → pass messages and status straight through, translate sendMessage/stop to onSend({ content })/onStop.

Registry reference

  • Index: https://agent-elements.21st.dev/r/index.json
  • Per-component: https://agent-elements.21st.dev/r/<id>.json
  • Full docs in one file: https://agent-elements.21st.dev/llms-full.txt

Related skills

vercel-react-best-practicesvercel-labs751KReact and Next.js performance optimization guidelines from Vercel Engineering. This skill should be used when writing, reviewing, or refactoring React/Next.js code to ensure optimal performance patterns. Triggers on tasks involving React components, Next.js pages, data fetching, bundle optimization, or performance improvements.vercel-composition-patternsvercel-labs361KReact composition patterns that scale. Use when refactoring components with boolean prop proliferation, building flexible component libraries, or designing reusable APIs. Triggers on tasks involving compound components, render props, context providers, or component architecture. Includes React 19 API changes.vercel-react-view-transitionsvercel-labs135KGuide for implementing smooth, native-feeling animations using React's View Transition API (`<ViewTransition>` component, `addTransitionType`, and CSS view transition pseudo-elements). Use this skill whenever the user wants to add page transitions, animate route changes, create shared element animations, animate enter/exit of components, animate list reorder, implement directional (forward/back) navigation animations, or integrate view transitions in Next.js. Also use when the user mentions viewpick-ui-libraryemilkowalski124KPick the right library for a given frontend task from a curated, opinionated list — numbers, OTP inputs, charts, command menus, virtualization, drag and drop, toasts, state, styling, and more. Only runs when explicitly invoked; it does not trigger on its own.

Search skills and MCP servers

Fuzzy search across 23,137 skills and servers