Frontend Tools

Let your agent interact with and update your React Native app's UI.

What is this?

A frontend tool is a function that you register in your app. Your agent can call it, and the function runs on the device. Because the handler runs in your app, it has direct access to component state, native modules, and any library the app already uses. That's how an agent can "reach into" the app: update React state, switch a screen, open a sheet, or read local storage.

This page covers the "agent drives the UI" shape of frontend tools. To render a component for a tool call, see Components as Tools. To ask the user before a tool runs, see Human-in-the-loop.

When should I use this?

Use frontend tools when your agent needs to:

  • Read or modify React component state
  • Call device APIs through native modules, such as haptics, the clipboard, or storage
  • Navigate between screens
  • Show alerts, toasts, or bottom sheets
  • Interact with third-party libraries in your app

How it works in code

Nothing to wire on the agent

On every run, the OpenAI Agents AG-UI adapter adds each tool the frontend registered to a per-request copy of the agent, so the agent declares no tools of its own for this. When the model calls one, the run stops there, because the browser owns that call. The browser runs the handler, and the next request carries the result as a tool message, which the model sees as the call's output before it continues.

If a frontend tool has the same name as one of the agent's own tools, the agent's tool wins and the frontend one is ignored.

Tell the model when to call it

This is the part that is easy to miss. The tool arrives on every run, but a model with no instruction about it will answer in prose and never call it. Name the tool in the instructions and say what it is for.

gen_ui_tool_based.py
from agents import Agent

from demo_agents._common import MODEL, chat_settings

SYSTEM_PROMPT = """You are a data visualization assistant.

When the user asks for a chart, call `render_bar_chart` or `render_pie_chart`
with a concise title, short description, and a `data` array of
`{label, value}` items. Pick bar for comparisons over a small set of
categories; pick pie for composition / share-of-whole.

If the user names a chart subject but does NOT supply concrete numbers
(e.g. "show me a pie chart of website traffic by source"), do NOT ask
them for data. Invent plausible illustrative sample values yourself,
call the appropriate `render_*` tool immediately, and briefly note in
the follow-up that the values are illustrative samples. Always render
the chart on the first turn -- never reply with a clarifying question
asking for the data.

Keep chat responses brief -- let the chart do the talking."""


gen_ui_tool_based_agent = Agent(
    name="chart_assistant",
    instructions=SYSTEM_PROMPT,
    model=MODEL,
    model_settings=chat_settings(),
)

Backend tools in the same turn

A frontend tool call ends the run as soon as the model makes it. A backend tool the model called in the same turn and that has not finished is cancelled, and the next request answers it as not having run. Where that matters, turn off parallel tool calls on the agent with ModelSettings(parallel_tool_calls=False).

Register a frontend tool with useFrontendTool. Give it a name, a schema for its parameters, and a handler. The agent can then call it like any other tool, and your app runs it on the device.

parameters takes any Standard Schema, such as a zod schema. zod is an optional peer dependency of @copilotkit/react-native, so add it to your app:

npm install zod
src/TaskScreen.tsx
import { useState } from "react";
import { Text, View } from "react-native";
import { useFrontendTool } from "@copilotkit/react-native/headless";
import { z } from "zod";

export function TaskScreen() {
  const [filter, setFilter] = useState<"all" | "open" | "done">("all");

  useFrontendTool({
    name: "filter_tasks",
    description: "Filter the visible task list by status.",
    parameters: z.object({
      status: z.enum(["all", "open", "done"]),
    }),
    handler: async ({ status }) => {
      setFilter(status);
      return { applied: true, status };
    },
  });

  return (
    <View style={{ padding: 16 }}>
      <Text>Current filter: {filter}</Text>
    </View>
  );
}

The handler receives the parsed, type-safe parameters and can do anything your app can do: update state, call an API, or call a native module. Its return value is sent back to the agent as the tool result, so the model can reason about what happened.

Mount the component under CopilotKitProvider, next to your chat. This example uses the prebuilt chat from @copilotkit/react-native/components. A chat you build yourself, like the one in the quickstart, works the same way.

App.tsx
import { View } from "react-native";
import { CopilotKitProvider } from "@copilotkit/react-native/headless";
import { CopilotChat } from "@copilotkit/react-native/components";
import { TaskScreen } from "./src/TaskScreen";

// See the quickstart for emulator and device hosts.
const runtimeUrl = "http://localhost:8200/api/copilotkit";

export default function App() {
  return (
    <CopilotKitProvider runtimeUrl={runtimeUrl}>
      <View style={{ flex: 1 }}>
        <TaskScreen />
        <CopilotChat />
      </View>
    </CopilotKitProvider>
  );
}
Peer dependencies of the prebuilt chat

@copilotkit/react-native/components needs @gorhom/bottom-sheet, react-native-gesture-handler, react-native-reanimated, and react-native-streamdown. If you only use the hooks, import from /headless and skip them. See Import surfaces.

Ask the agent to show only open tasks. "Current filter" changes to open, and the agent replies after it reads the { applied: true } result.

Handlers read the values from registration

useFrontendTool registers the tool once and keeps that handler. If the handler reads a value from props or state, it sees the value from the last registration. Pass those values as the second argument, so the tool registers again when they change:

useFrontendTool(
  {
    name: "add_task",
    description: "Add a task to the list that is open on screen.",
    parameters: z.object({ title: z.string() }),
    handler: async ({ title }) => {
      await saveTask(listId, title); // your own storage call
      return { added: title, listId };
    },
  },
  [listId],
);

A state setter from useState never changes, so the first example needs no second argument.

Register one tool per call

useFrontendTool registers one tool per call. The React package also has useFrontendTools for a list of tools that changes between renders. @copilotkit/react-native does not export it, so call useFrontendTool once per tool, with the same number of calls on every render.

To hide a tool from the agent without removing it, set available: false. To offer a tool to one agent only, set agentId.

Draw the call in the chat

Add render to the same useFrontendTool call to show UI for the call in the chat. The prebuilt chat draws it, and so does a chat you build with useRenderToolCall. render must return a React Native element or null. A bare string type-checks, then throws inside the chat's FlatList. See Tool Call Rendering for the full pattern.

Before you ship

Your agent discovers the name, description, and parameter schema through the runtime connection. Before you ship a tool:

  • Keep the description narrow.
  • Validate the inputs.
  • Ask the user before an action that cannot be undone. Use useHumanInTheLoop for that.
  • Keep privileged changes, secret API calls, and authorization checks on your backend.

For every option, see the useFrontendTool reference.

How is this guide?