Tool UI

Tool UI

Use this skill to move from request to working Tool UI integration quickly.

Prefer assistant-ui when the project has no existing chat UI/runtime. Treat assistant-ui as optional when the app already has a working runtime.

Step 1: Compatibility and Doctor

Read components.json in the user's project and verify:

• components.json exists.

Step 2: Install Components

Install command from project root

Preferred (AI-assisted integration):

npx tool-agent "integrate the <component> component"

Use component-specific prompts from the catalog below (e.g. integrate the plan component for step-by-step task workflows with status tracking).

Alternative (direct registry):

npx shadcn@latest add @tool-ui/<component-id>

Multiple components:

npx tool-agent "integrate the plan, progress-tracker, and approval-card components for planning flows"

Or via shadcn:

npx shadcn@latest add @tool-ui/plan @tool-ui/progress-tracker @tool-ui/approval-card

Complete component catalog

All 25 Tool UI components with tool-agent prompts and shadcn install commands:

Progress

Input

Display

Artifacts

Confirmation

Media

Display (Geo Map)

Example installs by use case

tool-agent (recommended):

npx tool-agent "integrate the plan, progress-tracker, and approval-card components for planning flows"
npx tool-agent "integrate citation, link-preview, code-block, and code-diff for research output"
npx tool-agent "integrate data-table, chart, and stats-display for data visualization"
npx tool-agent "integrate image, image-gallery, video, and audio for media display"

shadcn (direct):

npx shadcn@latest add @tool-ui/plan @tool-ui/progress-tracker @tool-ui/approval-card
npx shadcn@latest add @tool-ui/citation @tool-ui/link-preview @tool-ui/code-block @tool-ui/code-diff
npx shadcn@latest add @tool-ui/data-table @tool-ui/chart @tool-ui/stats-display
# npm i recharts  # peer for chart
npx shadcn@latest add @tool-ui/image @tool-ui/image-gallery @tool-ui/video @tool-ui/audio

Toolkit setup in a codebase

After installing components, wire them into assistant-ui via a Toolkit. This section covers the full setup: provider, runtime, toolkit file, and ID handling.

1. Provider and runtime

Create an assistant wrapper that provides runtime, transport, and tools:

"use client";

import { lastAssistantMessageIsCompleteWithToolCalls } from "ai";
import { AssistantRuntimeProvider, Tools, useAui } from "@assistant-ui/react";
import {
  AssistantChatTransport,
  useChatRuntime,
} from "@assistant-ui/react-ai-sdk";
import { Thread } from "@/components/assistant-ui/thread";
import { toolkit } from "@/components/toolkit";

export const Assistant = () => {
  const runtime = useChatRuntime({
    transport: new AssistantChatTransport({ api: "/api/chat" }),
    sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls,
  });
  const aui = useAui({ tools: Tools({ toolkit }) });

  return (
    <AssistantRuntimeProvider runtime={runtime} aui={aui}>
      <div className="h-dvh">
        <Thread />
      </div>
    </AssistantRuntimeProvider>
  );
};

Key points:

• useChatRuntime + AssistantChatTransport: connects to your chat API.
• Tools({ toolkit }): forwards tool definitions and renderers to the model.
• sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls: auto-continues after tool calls (optional but common for tool-heavy flows).

2. Toolkit file structure

Create a single toolkit.ts (or toolkit.tsx) that exports a Toolkit object. Each key is a tool name; each value has type, description, parameters, and render.

Tool descriptions — Always include a description on every tool. Describe when to call the tool and what role it plays, not the visible content. The Tool UI component already renders the payload (options, plan, chart, etc.) to the user; a description that repeats that content is redundant. Prefer model-oriented guidance (e.g. "Present a plan for the user to follow" or "Let the user pick one option") over content echo (e.g. "Shows a list of options with labels and descriptions").

-Frontend vs backend tools

Backend tools (model returns result; no user input):

import { type Toolkit } from "@assistant-ui/react";
import { Plan } from "@/components/tool-ui/plan";
import { safeParseSerializablePlan } from "@/components/tool-ui/plan/schema";

export const toolkit: Toolkit = {
  showPlan: {
    type: "backend",
    render: ({ result }) => {
      const parsed = safeParseSerializablePlan(result);
      if (!parsed) return null;
      return <Plan {...parsed} />;
    },
  },
};

Frontend tools (model sends args; user interaction commits via addResult):

import { type Toolkit } from "@assistant-ui/react";
import { OptionList } from "@/components/tool-ui/option-list";
import {
  SerializableOptionListSchema,
  safeParseSerializableOptionList,
} from "@/components/tool-ui/option-list/schema";

const optionListTool: Toolkit[string] = {
  description: "Render selectable options with confirm and clear actions.",
  parameters: SerializableOptionListSchema,
  render: ({ args, toolCallId, result, addResult }) => {
    const parsed = safeParseSerializableOptionList({
      ...args,
      id: args?.id ?? `option-list-${toolCallId}`,
    });
    if (!parsed) return null;

    if (result) {
      return <OptionList {...parsed} choice={result} />;
    }
    return (
      <OptionList
        {...parsed}
        (actionId, selection) => {
          if (actionId === "confirm" || actionId === "cancel") {
            await addResult?.(selection);
          }
        }}
      />
    );
  },
};

export const toolkit: Toolkit = {
  option_list: optionListTool,
  approval_card: {
    /* ... */
  },
};

3. API route (AI SDK)

When the chat API uses the AI SDK (streamText), define backend tools with tool() from ai:

• Use inputSchema
• Backend tools use execute on the server; the result is streamed and rendered via the toolkit render function

import { streamText, tool, convertToModelMessages } from "ai";
import { openai } from "@ai-sdk/openai";
import { z } from "zod";

// With frontend tools: ...frontendTools(clientTools) — clientTools come from the request body via AssistantChatTransport
const result = streamText({
  model: openai("gpt-4o"),
  messages: await convertToModelMessages(messages),
  tools: {
    get_weather: tool({
      description:
        "Get the current weather and forecast for a location. Returns data to display in a weather widget.",
      inputSchema: z.object({
        location: z.string().describe("City name, e.g. 'San Francisco'"),
        units: z
          .enum(["celsius", "fahrenheit"])
          .default("fahrenheit")
          .describe("Temperature unit"),
      }),
      execute: async ({ location, units }) => {
        // Fetch weather data, return shape matching your widget schema
        return { location, units /* ... */ };
      },
    }),
  },
});

4. Action-centric vs compound components

Action-centric example (OptionList):

return (
  <OptionList
    {...parsed}
    (actionId, selection) => {
      if (actionId === "confirm" || actionId === "cancel") {
        await addResult?.(selection);
      }
    }}
  />
);

Compound example (OrderSummary with DecisionActions):

return (
  <ToolUI id={parsed.id}>
    <ToolUI.Surface>
      <OrderSummary {...parsed} />
    </ToolUI.Surface>
    <ToolUI.Actions>
      <ToolUI.DecisionActions
        actions={[
          { id: "cancel", label: "Cancel", variant: "outline" },
          { id: "confirm", label: "Purchase" },
        ]}
        =>
          createDecisionResult({ decisionId: parsed.id, action })
        }
        => addResult?.(decision)}
      />
    </ToolUI.Actions>
  </ToolUI>
);

ApprovalCard uses embedded actions; wire onConfirm/onCancel directly:

return (
  <ApprovalCard
    {...parsed}
    choice={
      result === "approved" || result === "denied" ? result : parsed.choice
    }
    () => addResult?.("approved")}
    () => addResult?.("denied")}
  />
);

Action Model

Tool UI uses two action surfaces, rendered as compound siblings outside the display component:

• ToolUI.LocalActions: non-consequential side effects (export, copy, open link). Handlers must not call addResult(...).
• ToolUI.DecisionActions: consequential choices that produce a DecisionResult envelope via createDecisionResult(...). The commit callback calls addResult(...).

Compound wrapper pattern for display components with actions:

<ToolUI id={surfaceId}>
  <ToolUI.Surface>
    <DataTable {...props} />
  </ToolUI.Surface>
  <ToolUI.Actions>
    <ToolUI.LocalActions
      actions={[{ id: "export-csv", label: "Export CSV" }]}
      => {
        /* side effects only */
      }}
    />
  </ToolUI.Actions>
</ToolUI>

Three components are action-centric exceptions — they keep embedded action props instead of sibling surfaces. All three share a unified interface:

• actions: action buttons rendered by the component.
• onAction(actionId, state): runs after the action and receives post-action state.
• onBeforeAction(actionId, state): guard evaluated before an action runs.

Components using the compound pattern: CodeBlock, CodeDiff, Terminal, ProgressTracker.

Context is shared via createContext + use() (React 19). Subcomponents throw if used outside their Root.

Receipt and Choice Convention

Components with outcomes use a choice prop to render confirmed/completed state:

When choice is present, the component renders in receipt mode — read-only, no actions.

Operational Rules

• Install the smallest set of components that solves the request.
• Every tool must have a description; write it for the model (when to call, what role) — avoid repeating content the component will render.

Notes: Frontend tools need an execute function Backend tools have the tool implementation on the server side. Backend tool don't need either Ignore the generated files After setup:

• Ensure the required package dependencies are installed so the first experience of running after the changes is magical
• Notify the user if env variables are not set that should be for a successful run of the feature that was just implemented, most likely mainly variables required by the api chat endpoint.