Tool Call Rendering

Render your agent's tool calls with your own React Native components.

What is this?

Tools are how an LLM invokes predefined, typically deterministic functions. Tool rendering lets you decide how each of those tool calls appears in the chat. You register a React Native component that draws a card for the call: its arguments, a live status, and the result once it arrives. This is the generative UI variant CopilotKit calls tool rendering.

When should I use this?

Render tool calls when you want to:

  • Show users which tools the agent is calling, and with what arguments
  • Display live progress while a tool runs
  • Render rich results once a tool completes
  • Give tool-heavy agents a clear, on-brand chat

Default rendering on React Native

The React package has useDefaultRenderTool, which draws a built-in status card for every tool call. Its card uses web elements, so @copilotkit/react-native does not export it.

Without a renderer, the rendered CopilotChat from /components shows a short "Called: tool_name" line for each call. A chat you build yourself shows nothing, unless you render a fallback. To draw every call your way, register a catch-all.

Custom catch-all

useRenderTool with name: "*" registers one renderer for every tool call that has no renderer of its own:

src/ToolCallCard.tsx
import { ActivityIndicator, Text, View } from "react-native";
import { useRenderTool } from "@copilotkit/react-native/headless";

export function useToolCallCards() {
  useRenderTool(
    {
      name: "*",
      render: ({ name, status, result }) => (
        <View style={{ padding: 10, borderRadius: 10, backgroundColor: "#f4f4f8", gap: 4 }}>
          <View style={{ flexDirection: "row", alignItems: "center", gap: 8 }}>
            {status === "complete" ? <Text>✓</Text> : <ActivityIndicator size="small" />}
            <Text style={{ fontWeight: "600" }}>{name}</Text>
          </View>
          {status === "complete" ? <Text numberOfLines={3}>{result}</Text> : null}
        </View>
      ),
    },
    [],
  );
}

Call useToolCallCards() once, in a component under CopilotKitProvider. Every tool call without a named renderer now draws this card.

Per-tool renderers

The most expressive path is one renderer per tool name. Each renderer receives the tool's parsed parameters, a live status, and, once the agent returns, the result. The renderer must return a React Native element or null.

src/WeatherCard.tsx
import { Text, View } from "react-native";
import { useRenderTool } from "@copilotkit/react-native/headless";
import { z } from "zod";

export function useWeatherCard() {
  useRenderTool(
    {
      name: "get_weather", // must match the tool name the agent exposes
      parameters: z.object({ location: z.string() }),
      render: ({ parameters, status, result }) => {
        if (status !== "complete") {
          return <Text>Checking the weather in {parameters.location ?? "..."}</Text>;
        }
        const weather = parseWeather(result);
        return (
          <View style={{ padding: 12, borderRadius: 12, backgroundColor: "#eef4ff" }}>
            <Text style={{ fontWeight: "600" }}>{parameters.location}</Text>
            <Text>{weather ? `${weather.temperature}°` : result}</Text>
          </View>
        );
      },
    },
    [],
  );
}

// The result arrives as a string. Parse it when your tool returns JSON.
function parseWeather(result: string): { temperature: number } | null {
  try {
    return JSON.parse(result);
  } catch {
    return null;
  }
}

Named renderers and a catch-all work together: the named one claims its tool, and the "*" renderer handles every other call.

The name you pass to useRenderTool must match the tool name the agent exposes. That is how the chat matches the call to your component.

Tool inputs and results are separate

In useRenderTool, parameters holds the inputs the agent sent to the tool. It does not turn into the tool's return value when status becomes "complete". The output arrives separately as result, a string. For a tool that returns JSON, parse that string before you read its fields.

While the arguments stream, status is "inProgress" and parameters can be partial. Read required fields after the status moves to "executing" or "complete".

Your own tool or someone else's

useRenderTool registers a renderer only. It takes no description and no handler, so the model never sees it as a tool. Use it for a tool that runs on your agent's backend, or as the "*" fallback. For a tool your app owns and runs on the device, pass render to useFrontendTool instead. Both write to the same renderer registry. If your code passes description or handler to useRenderTool, see Migrating from the old React Native useRenderTool.

Where the cards render

The rendered CopilotChat and CopilotModal from /components draw tool calls with your renderers. In your own chat, useRenderToolCall() returns a function that renders one tool call. It returns ReactElement | null, which is what a FlatList renderItem needs. See Headless UI.

The backend tool definition

The frontend renderer only sees what the agent sends. The matching backend must expose a tool named get_weather and return structured data. The renderer with the same name then draws the card. The React Native renderer does not create or run that backend tool.

tool_rendering_agent.py
from random import choice, randint

from langchain.agents import create_agent
from langchain.tools import tool
from langchain_openai import ChatOpenAI
from copilotkit import CopilotKitMiddleware

# Multi-tool chaining prompt.
#
# The goal of this demo is to surface MULTIPLE tool-call cards per turn so
# the rendering patterns (per-tool + catch-all) get exercised visibly. The
# prompt nudges the model toward an explore-then-enrich pattern (e.g.
# `get_weather("Tokyo")` -> `search_flights(..., "Tokyo")`) without forcing
# a rigid recipe: we describe the *habit*, not a chain.
SYSTEM_PROMPT = (
    "You are a helpful travel & lifestyle concierge. You have mock tools "
    "for weather, flights, stock prices, and dice rolls - they all return "
    "fake data, so call them liberally.\n\n"
    "Your habit is to CHAIN tools when one answer naturally invites another. "
    "For a single user question, call at least TWO tools in succession when "
    "the topic allows before composing your final reply. Examples of "
    "helpful chains you should default to:\n"
    "  - 'What's the weather in Tokyo?' -> call get_weather('Tokyo'), then "
    "call search_flights(origin='SFO', destination='Tokyo') so the user "
    "also sees how to get there.\n"
    "  - 'How is AAPL doing?' -> call get_stock_price('AAPL'), then call "
    "get_stock_price on a related ticker (e.g. 'MSFT' or 'GOOGL') for "
    "comparison.\n"
    "  - 'Roll a d20' -> call roll_dice(20), then call roll_dice again with "
    "a different number of sides so the user sees a contrast.\n"
    "  - 'Find flights from SFO to JFK' -> call search_flights, then call "
    "get_weather on the destination city.\n\n"
    "Only skip chaining when the user has clearly asked for a single, "
    "atomic answer and more tool calls would feel intrusive. Never "
    "fabricate data that a tool could provide."
)


@tool
def get_weather(location: str) -> dict:
    """Get the current weather for a given location.

    Useful on its own for weather questions, and a great companion to
    `search_flights` - always consider checking the weather at a
    destination the user is flying to, and checking flights to any
    city whose weather the user has just asked about.
    """
    return {
        "city": location,
        "temperature": 68,
        "humidity": 55,
        "wind_speed": 10,
        "conditions": "Sunny",
    }

To check it, ask the agent for the weather in San Francisco. Your card renders with the temperature from result.

For every option, see the useRenderTool reference.

How is this guide?