assistant-ui Tools

Always consult assistant-ui.com/llms.txt for the latest API.

Tools let LLMs trigger actions with custom UI rendering.

References

• ./references/make-tool.md -- makeAssistantTool/useAssistantTool
• ./references/tool-ui.md -- makeAssistantToolUI rendering
• ./references/human-in-loop.md -- Confirmation patterns
• ./references/toolkits.md -- Toolkits and the Tools component
• ./references/mcp-server.md -- Server-side MCP tools
• ./references/generative-ui.md -- Declarative generative UI
• ./references/registry-components.md -- ToolFallback, ToolGroup, Image renderers

Authoring Model

Toolkits are the supported path:

"use generative" file exporting defineToolkit({ ... })
  → registered with useAui({ tools: Tools({ toolkit }) })
  → exposed to the model with new AISDKToolkit({ toolkit })

The component and hook APIs are deprecated: makeAssistantTool, useAssistantTool, makeAssistantToolUI, and useAssistantToolUI all carry @deprecated and point at the toolkit migration guide. They still work, so existing code is not broken, but new tools should be toolkit entries. For per-message UI that used to need makeAssistantToolUI, use inline tool render overrides on MessagePrimitive.Parts.

See ./references/toolkits.md for the full authoring guide.

Toolkit (recommended)

One file holds the schema, the executor, and the renderer. A build plugin splits it so a backend execute never ships to the browser and a render never reaches the server.

// app/toolkit.tsx
"use generative";

import { defineToolkit } from "@assistant-ui/react";
import { z } from "zod";

export default defineToolkit({
  get_weather: {
    description: "Get current weather for a location.",
    parameters: z.object({ location: z.string() }),
    execute: async ({ location }) => fetchWeatherAPI(location),
    render: ({ args, result }) =>
      result ? <div>{result.temperature}°</div> : <div>Fetching {args.location}…</div>,
  },
});

Mount it through useAui, and expose the same import to the model on the server:

const runtime = useChatRuntime();
const aui = useAui({ tools: Tools({ toolkit }) });

<AssistantRuntimeProvider aui={aui} runtime={runtime}>{children}</AssistantRuntimeProvider>

// app/api/chat/route.ts
const aiToolkit = new AISDKToolkit({ toolkit });
const result = streamText({
  model: openai("gpt-5.4-nano"),
  messages: await convertToModelMessages(messages),
  tools: await aiToolkit.tools({ frontend: tools }),
});

The directive requires the compiler: withAui() from @assistant-ui/next, aui() from @assistant-ui/vite, or withAui() from @assistant-ui/metro.

Backend Tool with UI (component model)

// Backend (app/api/chat/route.ts)
import { openai } from "@ai-sdk/openai";
import {
  streamText,
  tool,
  zodSchema,
  stepCountIs,
  convertToModelMessages,
  createUIMessageStreamResponse,
  toUIMessageStream,
  type UIMessage,
} from "ai";
import { z } from "zod";

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: openai("gpt-5.4-nano"),
    messages: await convertToModelMessages(messages),
    stopWhen: stepCountIs(5),
    tools: {
      get_weather: tool({
        description: "Get weather for a city",
        inputSchema: zodSchema(z.object({ city: z.string() })),
        execute: async ({ city }) => ({ temp: 22, city }),
      }),
    },
  });

  return createUIMessageStreamResponse({
    stream: toUIMessageStream({ stream: result.stream }),
  });
}

import { makeAssistantToolUI } from "@assistant-ui/react";

const WeatherToolUI = makeAssistantToolUI({
  toolName: "get_weather",
  render: ({ args, result, status }) => {
    // status is an object; check status.type (not status === "running")
    if (status.type === "running") return <div>Loading weather...</div>;
    return <div>{result?.city}: {result?.temp}°C</div>;
  },
});

<AssistantRuntimeProvider runtime={runtime}>
  <WeatherToolUI />
  <Thread />
</AssistantRuntimeProvider>

Frontend-Only Tool

import { makeAssistantTool } from "@assistant-ui/react";
import { z } from "zod";

const CopyTool = makeAssistantTool({
  toolName: "copy_to_clipboard",
  parameters: z.object({ text: z.string() }),
  execute: async ({ text }) => {
    await navigator.clipboard.writeText(text);
    return { success: true };
  },
});

<AssistantRuntimeProvider runtime={runtime}>
  <CopyTool />
  <Thread />
</AssistantRuntimeProvider>

API Reference

// makeAssistantToolUI render props (ToolCallMessagePartProps)
interface ToolCallMessagePartProps {
  toolCallId: string;
  toolName: string;
  args: Record<string, unknown>;
  argsText: string;          // raw streamed JSON
  result?: unknown;
  isError?: boolean;
  artifact?: unknown;        // UI-only artifact attached to the result

  // status is an OBJECT, not a string. Branch on status.type.
  status:
    | { type: "running" }
    | { type: "complete" }
    | { type: "incomplete"; reason: "cancelled" | "length" | "content-filter" | "other" | "error" }
    | { type: "requires-action"; reason: "interrupt" };

  // Supply a result from the renderer (instead of a tool execute function)
  addResult: (result: unknown | ToolResponse<unknown>) => void;
  // Resume a tool paused for human input
  resume: (payload: unknown) => void;
  // Respond to a server-side approval gate; reads the approval id off the part
  respondToApproval: (response: ToolApprovalResponse) => void;
}

// One of:
type ToolApprovalResponse =
  | { approved: boolean; reason?: string }
  | { optionId: string; reason?: string }
  | { approved: boolean; optionId: string; reason?: string };

Render props also carry the enriched part state, so toolUI, argsText, and the part status are all available without a second lookup.

Server-Side Approval Gates

AI SDK v7 can pause a tool call server-side two ways, and they coexist: needsApproval on the tool definition gates that tool everywhere, and the call-level toolApproval option gates per call so the decision can vary by input. Either way assistant-ui surfaces the gate as approval on the tool part, and respondToApproval is the only correct way to acknowledge it.

// Backend
const result = streamText({
  model: openai("gpt-5.4-nano"),
  messages: await convertToModelMessages(messages),
  tools: { deploy: tool({ /* ... */ }) },
  toolApproval: {
    deploy: (input) =>
      input.target === "production" ? "user-approval" : "not-applicable",
  },
});

// Frontend: post the decision back automatically
import { lastAssistantMessageIsCompleteWithApprovalResponses } from "ai";

const runtime = useChatRuntime({
  sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithApprovalResponses,
});

// Renderer
render: ({ args, approval, respondToApproval }) => {
  if (approval?.approved === undefined) {
    return (
      <div>
        <p>Approve deploy to {args.target}?</p>
        <button onClick={() => respondToApproval({ approved: true })}>Approve</button>
        <button onClick={() => respondToApproval({ approved: false, reason: "denied" })}>
          Deny
        </button>
      </div>
    );
  }
  return approval.approved ? <p>Approved</p> : <p>Denied</p>;
};

Human-in-the-Loop

const DeleteToolUI = makeAssistantToolUI({
  toolName: "delete_file",
  render: ({ args, status, addResult }) => {
    if (status.type === "requires-action") {
      return (
        <div>
          <p>Delete {args.path}?</p>
          <button onClick={() => addResult({ confirmed: true })}>Confirm</button>
          <button onClick={() => addResult({ confirmed: false })}>Cancel</button>
        </div>
      );
    }
    return <div>File deleted</div>;
  },
});

Common Gotchas

Tool UI not rendering

• toolName must match exactly (case-sensitive); in a toolkit the object key is the tool name
• Register UI inside AssistantRuntimeProvider

"use generative" file has no effect

• The directive needs a compiler. Add withAui() (@assistant-ui/next), aui() (@assistant-ui/vite), or withAui() (@assistant-ui/metro)

Tool not being called

• Check tool description is clear
• Use stopWhen: stepCountIs(n) to allow multi-step

Result not showing

• Tool must return a value
• Check status.type === "complete" before accessing result

Reading the tool UI registry

• s.tools.tools was removed in 0.15. Select s.tools.toolUIs[toolName]?.[0]?.render; entries carry the renderer alongside its presentation options