CopilotKit

CopilotKit provides a full React chat runtime and pairs especially well with LangGraph when you want the agent to return structured UI payloads instead of only plain text. The frontend talks to a CopilotKit runtime URL and parses assistant messages into dynamic React components.

In this pattern, your LangGraph deployment serves the graph API and an AG-UI FastAPI bridge. A small Node CopilotKit runtime sits in front of that bridge and is what the React client calls.

On the server, the copilotkit package provides CopilotKitMiddleware so a LangGraph graph, a LangChain agent, or a Deep Agent can speak the Agent UI (AG-UI) wire protocol, stream tool and message events to a chat UI, and read or write the shared CopilotKit slice of state. TypeScript setups usually mount the CopilotKit runtime next to the graph API. Python setups mount an AG-UI FastAPI bridge on the deployment and run the CopilotKit runtime in Node in front of it.

This approach is useful when you want:

• a ready-made chat runtime instead of wiring stream.messages yourself
• a custom server endpoint that can add provider-specific behavior next to your deployed graph
• structured generative UI rendered from a constrained component registry

CopilotKit for LangGraph also documents generative UI, human in the loop (HITL), and shared state on top of the same middleware and clients.

Info

For CopilotKit-specific APIs, UI patterns, and runtime configuration, see the CopilotKit docs. For a Deep Agent walkthrough, see Deep Agents and CopilotKit in the CopilotKit docs.

Interactive example: ExampleEmbed

How it works

At a high level, CopilotKit sits between your React app and the LangGraph deployment.

The frontend sends conversation state to a Node CopilotKit runtime. That runtime calls the AG-UI FastAPI bridge on your LangGraph deployment, and the response comes back with both assistant messages and any structured UI payloads your component registry can render.

1. Deploy the graph as usual using LangSmith or using a LangGraph development server.
2. Extend the deployment with an HTTP app that mounts the AG-UI FastAPI bridge next to the graph API.
3. Run a Node CopilotKit runtime that points an HttpAgent at that bridge, then wrap the frontend in CopilotKit against the runtime URL.
4. Register dynamic UI components and parse assistant responses into those components at render time.

    %%{
  init: {
    "fontFamily": "monospace",
    "flowchart": {
      "curve": "curve"
    }
  }
}%%
graph LR
  USER["User input"]
  UI["CopilotKit React app"]
  ENDPOINT["/api/copilotkit"]
  GRAPH["LangGraph deployment"]
  RENDER["Hashbrown UI kit"]

  USER --> UI
  UI --> RUNTIME
  RUNTIME --> GRAPH
  GRAPH --> RUNTIME
  RUNTIME --> UI
  UI --> RENDER
  

On Python, the CopilotKit runtime is a separate Node process. The LangGraph deployment hosts the graph API and the AG-UI FastAPI bridge only. See Add the CopilotKit runtime.

What you get on the Python server

The copilotkit and related packages bridge a LangGraph deployment and CopilotKit clients.

CopilotKitMiddleware is the same middleware for create_deep_agent and for a graph from create_agent when you add it to the middleware list. For a create_agent graph with CopilotKitState and a FastAPI bridge, follow the Python main.py example below. Structured generative UI (for example useAgentContext and an output_schema from the client) needs extra middleware that maps Copilot state to a structured output strategy, as in the expandable src/middleware.py example in the same section.

Mounting app on the http key in langgraph.json follows the usual LangGraph or LangSmith deployment: one process serves the graph API and the AG-UI FastAPI bridge. The CopilotKit React client talks to the Node runtime, which calls that bridge.

Installation

For the backend endpoint:

uv add copilotkit ag-ui-langgraph fastapi uvicorn

The middleware package sits alongside the Deep Agents stack. Install it with your chat model package (this example uses OpenAI):

pip

pip install -U deepagents copilotkit langchain-openai

uv

uv add deepagents copilotkit langchain-openai

The CopilotKit runtime that the frontend calls is a Node package, so the Python setup also needs a small Node server. The pins below are a known-good pair. If you upgrade @copilotkit/runtime, install the @ag-ui/client version that release lists as a dependency so the HttpAgent type matches:

bun add @copilotkit/runtime@1.73.3 @ag-ui/client@0.0.59

For the frontend app:

bun add @copilotkit/react-core @copilotkit/react-ui @hashbrownai/core @hashbrownai/react

Use CopilotKit with a Deep Agent

Add CopilotKitMiddleware to the middleware list you pass to create_deep_agent. The middleware lets CopilotKit route frontend tool calls and align chat state with your graph. Keep any other middleware you configure in the same list.

The compiled graph is then ready to plug into a CopilotKit- or AG-UI–aware process (for example, the FastAPI pattern below) or a guide such as Deep Agents and CopilotKit in the CopilotKit documentation. Do not pass a checkpointer on the exported agent if you list it under graphs in langgraph.json. The FastAPI example below attaches a checkpointer only on the AG-UI route copy.

from deepagents import create_deep_agent
from copilotkit import CopilotKitMiddleware


def get_weather(location: str) -> str:
    """Return a simple weather string for a location."""
    return f"The weather in {location} is sunny."


agent = create_deep_agent(
    model="openai:gpt-5.5",
    tools=[get_weather],
    middleware=[CopilotKitMiddleware()],  # AG-UI, frontend tools, and context
    system_prompt="You are a helpful research assistant.",
)

Extend the LangGraph deployment with a custom endpoint

The key idea is that the LangGraph deployment does not only serve graphs. It can also load an HTTP app, which lets you mount extra routes next to the deployment itself.

In langgraph.json, point http.app at your custom app entrypoint:

{
  "dependencies": ["."],
  "graphs": {
    "copilotkit_shadify": "./main.py:agent"
  },
  "http": {
    "app": "./main.py:app"
  }
}

In Python, create a FastAPI app and expose the LangGraph agent through CopilotKit’s AG-UI bridge:

from typing import Any, TypedDict

from ag_ui_langgraph import add_langgraph_fastapi_endpoint
from copilotkit import CopilotKitMiddleware, CopilotKitState, LangGraphAGUIAgent
from fastapi import FastAPI
from langchain.agents import create_agent
from langgraph.checkpoint.memory import MemorySaver

from src.middleware import apply_structured_output_schema, normalize_context


class AgentState(CopilotKitState):
    pass


class AgentContext(TypedDict, total=False):
    output_schema: dict[str, Any]


agent = create_agent(
    model="openai:gpt-5.5",
    middleware=[
        normalize_context,
        CopilotKitMiddleware(),
        apply_structured_output_schema,
    ],
    context_schema=AgentContext,
    state_schema=AgentState,
    system_prompt=(
        "You are a helpful UI assistant. Build visual responses using the "
        "available components."
    ),
)

app = FastAPI()

add_langgraph_fastapi_endpoint(
    app=app,
    agent=LangGraphAGUIAgent(
        name="copilotkit_shadify",
        description="A UI assistant that returns structured component payloads.",
        # The AG-UI bridge reads graph state, so this route needs a checkpointer.
        graph=agent.copy(update={"checkpointer": MemorySaver()}),
    ),
    path="/",
)

The AG-UI bridge reads the graph’s state on every run, so the graph behind the FastAPI route needs a checkpointer. Keep the checkpointer off agent itself: agent is also listed under graphs, and langgraph dev refuses to load a listed graph that brings its own checkpointer, because LangGraph API handles persistence for listed graphs. The copy with MemorySaver serves only the FastAPI route. Its threads live in process memory, so they do not survive a restart, are not shared across replicas, and do not appear in the deployment’s threads API. For production traffic on this route, pass a durable checkpointer instead.

This custom app is the important extension point: it mounts the AG-UI FastAPI bridge without replacing the underlying LangGraph deployment. The Node CopilotKit runtime stays separate; see Add the CopilotKit runtime.

In Python, the equivalent work happens in middleware: normalize the CopilotKit context and forward the output_schema from useAgentContext(...) into the model’s structured output configuration. The schema that Hashbrown generates has no top-level title, and the OpenAI integration requires one to name the structured output schema, so the middleware fills in a title and description when they are missing.

import json
from collections.abc import Mapping

from langchain.agents.middleware import before_agent, wrap_model_call
from langchain.agents.structured_output import ProviderStrategy


@wrap_model_call
async def apply_structured_output_schema(request, handler):
    schema = None
    runtime = getattr(request, "runtime", None)
    runtime_context = getattr(runtime, "context", None)

    if isinstance(runtime_context, Mapping):
        schema = runtime_context.get("output_schema")

    if schema is None and isinstance(getattr(request, "state", None), dict):
        copilot_context = request.state.get("copilotkit", {}).get("context")
        if isinstance(copilot_context, list):
            for item in copilot_context:
                if isinstance(item, dict) and item.get("description") == "output_schema":
                    schema = item.get("value")
                    break

    if isinstance(schema, str):
        try:
            schema = json.loads(schema)
        except json.JSONDecodeError:
            schema = None

    if isinstance(schema, dict):
        # OpenAI uses the top-level title as the structured output name.
        schema = {
            **schema,
            "title": schema.get("title") or "CopilotKitStructuredOutput",
            "description": schema.get("description")
            or "Structured response schema for the CopilotKit preview.",
        }
        request = request.override(
            response_format=ProviderStrategy(schema=schema, strict=True),
        )

    return await handler(request)


@before_agent
def normalize_context(state, runtime):
    copilotkit_state = state.get("copilotkit", {})
    context = copilotkit_state.get("context")

    if isinstance(context, list):
        normalized = [
            item.model_dump() if hasattr(item, "model_dump") else item
            for item in context
        ]
        return {"copilotkit": {**copilotkit_state, "context": normalized}}

    return None

The result is a clean separation of concerns:

• LangGraph still owns graph execution and persistence
• CopilotKit owns the chat-facing runtime contract
• your custom endpoint glues them together inside one deployment

Add the CopilotKit runtime

The FastAPI route speaks the AG-UI protocol, not the CopilotKit runtime protocol, so the frontend cannot call it directly: pointing runtimeUrl at the FastAPI route fails with runtime_info_fetch_failed. Run a CopilotKit runtime in a small Node server and point an HttpAgent at the FastAPI route. The Python graph and middleware still define tool behavior and agent logic.

import { createServer } from "node:http";

import { HttpAgent } from "@ag-ui/client";
import { CopilotRuntime } from "@copilotkit/runtime/v2";
import { createCopilotNodeListener } from "@copilotkit/runtime/v2/node";

// The FastAPI route is mounted at the root of the LangGraph deployment.
const agentUrl = process.env.LANGGRAPH_DEPLOYMENT_URL || "http://127.0.0.1:2024";

const runtime = new CopilotRuntime({
  agents: {
    default: new HttpAgent({ url: agentUrl }),
  },
});

createServer(
  createCopilotNodeListener({
    runtime,
    basePath: "/api/copilotkit",
    cors: true,
  }),
).listen(4000);

Start the runtime next to the LangGraph deployment:

bun server.ts

The runtime serves /api/copilotkit on port 4000. Point the frontend at it, either by setting VITE_RUNTIME_URL=http://localhost:4000/api/copilotkit or by proxying /api/copilotkit to port 4000 from your dev server.

This runtime is the basic in-app integration. Channels does not use the FastAPI route: it connects to the graph through the deployment’s graph API with LangGraphAgent.

Structure the frontend app

On the frontend, wrap your app in CopilotKit and point it at the custom runtime URL:

import { CopilotKit } from "@copilotkit/react-core";
import { CopilotChat, useAgentContext } from "@copilotkit/react-core/v2";
import { s } from "@hashbrownai/core";

import { useChatKit } from "@/components/chat/chat-kit";
import { chatTheme } from "@/lib/chat-theme";

export function App() {
  return (
    <CopilotKit runtimeUrl={import.meta.env.VITE_RUNTIME_URL ?? "/api/copilotkit"}>
      <Page />
    </CopilotKit>
  );
}

function Page() {
  const chatKit = useChatKit();

  useAgentContext({
    description: "output_schema",
    value: s.toJsonSchema(chatKit.schema),
  });

  return <CopilotChat {...chatTheme} />;
}

There are two important pieces here:

• runtimeUrl="/api/copilotkit" sends the chat to your custom backend route rather than directly to the raw LangGraph API
• useAgentContext(...) sends the UI schema to the agent so the model knows what structured output format it should produce

Register the dynamic components

The component registry lives in useChatKit(). This is where you define the set of components the agent is allowed to emit, such as cards, rows, columns, charts, code blocks, and buttons.

import { s } from "@hashbrownai/core";
import { exposeComponent, exposeMarkdown, useUiKit } from "@hashbrownai/react";

import { Button } from "@/components/ui/button";
import { Card } from "@/components/ui/card";
import { CodeBlock } from "@/components/ui/code-block";
import { Row, Column } from "@/components/ui/layout";
import { SimpleChart } from "@/components/ui/simple-chart";

export function useChatKit() {
  return useUiKit({
    components: [
      exposeMarkdown(),
      exposeComponent(Card, {
        name: "card",
        description: "Card to wrap generative UI content.",
        children: "any",
      }),
      exposeComponent(Row, {
        name: "row",
        props: {
          gap: s.string("Tailwind gap size") as never,
        },
        children: "any",
      }),
      exposeComponent(Column, {
        name: "column",
        children: "any",
      }),
      exposeComponent(SimpleChart, {
        name: "chart",
        props: {
          labels: s.array("Category labels", s.string("A label")),
          values: s.array("Numeric values", s.number("A value")),
        },
        children: false,
      }),
      exposeComponent(CodeBlock, {
        name: "code_block",
        props: {
          code: s.streaming.string("The code to display"),
          language: s.string("Programming language") as never,
        },
        children: false,
      }),
      exposeComponent(Button, {
        name: "button",
        children: "text",
      }),
    ],
  });
}

This registry becomes the contract between the agent and the UI. The model is not generating arbitrary JSX. It is generating structured data that must validate against the components and props you exposed.

Render assistant messages as dynamic UI

Once the assistant response arrives, the custom message renderer decides how to display it. In this example:

• assistant messages are parsed as structured JSON against the UI kit schema
• valid structured output is rendered as real React components
• user messages are rendered as ordinary chat bubbles

import type { AssistantMessage } from "@ag-ui/core";
import type { RenderMessageProps } from "@copilotkit/react-ui";
import { useJsonParser } from "@hashbrownai/react";
import { memo } from "react";

import { useChatKit } from "@/components/chat/chat-kit";
import { Squircle } from "@/components/squircle";

const AssistantMessageRenderer = memo(function AssistantMessageRenderer({
  message,
}: {
  message: AssistantMessage;
}) {
  const kit = useChatKit();
  const { value } = useJsonParser(message.content ?? "", kit.schema);

  if (!value) return null;

  return (
    <div className="group/msg mt-2 flex w-full justify-start">
      <div className="magic-text-output w-full px-1 py-1">{kit.render(value)}</div>
    </div>
  );
});

export function CustomMessageRenderer({ message }: RenderMessageProps) {
  if (message.role === "assistant") {
    return <AssistantMessageRenderer message={message} />;
  }

  return (
    <div className="flex w-full justify-end">
      <Squircle className="w-full max-w-[64ch] px-4 py-3">
        <pre>{typeof message.content === "string" ? message.content : JSON.stringify(message.content, null, 2)}</pre>
      </Squircle>
    </div>
  );
}

This renderer pattern is what makes the integration feel native:

• CopilotKit handles chat state and transport
• the custom renderer decides how assistant payloads become UI
• Hashbrown turns validated structured data into concrete React elements

Channels

The same agent that powers your in-app copilot can also run as a bot in Slack and other messaging platforms. CopilotKit Channels connects your LangChain agent to a messaging platform through a managed CopilotKit Intelligence connection: Intelligence holds the platform credentials and message delivery, while your process runs the agent.

Note

Channels require @copilotkit/channels 0.6.1 and @copilotkit/runtime 1.65.0, installed together as a tested pair, and Node.js 22 or later on a long-running host. The LangChain deployment the channel connects to can be Python or TypeScript.

Info

This page shows how Channels fit together with a LangChain agent. For the full Slack walkthrough with a LangGraph backend, see Connect and run your agent in Slack. For other agent backends and platform coverage, see the Channels overview.

How it fits together

CopilotKit Intelligence owns the platform connection (Slack, Microsoft Teams, and more) and its credentials. Your long-running Channels listener owns the agent, tools, and application logic. Each turn follows the same path:

1. A person messages your app in Slack.
2. Intelligence receives the platform event using the credentials configured for the Channel.
3. A persistent Intelligence gateway connection delivers the turn to your Channels listener.
4. The listener runs your LangChain agent over AG-UI and renders the reply.
5. Intelligence sends the result back as native Slack Block Kit.

Platform credentials never enter the agent process. Create a managed Channel in CopilotKit Intelligence and connect Slack: Intelligence stores the Slack credentials and gives you a Channel code (CHANNEL_CODE) and a project-scoped Intelligence API key (INTELLIGENCE_API_KEY) for the listener.

Install

Install the tested SDK pair, then the TypeScript tooling:

npm

npm install --save-exact @copilotkit/channels@0.6.1 @copilotkit/runtime@1.65.0
npm install -D tsx typescript @types/node

pnpm

pnpm add --save-exact @copilotkit/channels@0.6.1 @copilotkit/runtime@1.65.0
pnpm add -D tsx typescript @types/node

yarn

yarn add --exact @copilotkit/channels@0.6.1 @copilotkit/runtime@1.65.0
yarn add -D tsx typescript @types/node

Use a NodeNext TypeScript configuration:

{
  "compilerOptions": {
    "target": "ES2022",
    "module": "NodeNext",
    "moduleResolution": "NodeNext",
    "strict": true,
    "skipLibCheck": true,
    "noEmit": true,
    "types": ["node"]
  },
  "include": ["*.ts", "*.tsx"]
}

Connect your LangChain agent

Return a fresh agent for each conversation, keyed by thread, so no state leaks across Slack threads. LangGraph Server speaks the LangGraph API, so use LangGraphAgent from @copilotkit/runtime/langgraph, not HttpAgent. Point it at the deployment you already run:

Run the Channel listener

Declare the Channel with the Code from Intelligence and forward each message to the agent. Creating the Node listener is what connects the Channel. Register process teardown before you create the listener so a Ctrl-C during the connect window still stops the Channel cleanly.

Set the runtime secrets, then start the process:

INTELLIGENCE_API_KEY=<project-api-key>
CHANNEL_CODE=support-slack
LANGGRAPH_DEPLOYMENT_URL=http://127.0.0.1:2024
PORT=3000

node --env-file=.env --import tsx channel.ts

The Channel moves from Waiting for runtime to Online in Intelligence once the process connects.

Other platforms

Managed Slack is generally available. Managed Teams is a controlled integration target. The Channels SDK also ships direct adapters for Teams, Discord, Telegram, and WhatsApp. See the Channels overview for the current platform list and per-platform setup.

Resources

• Deep Agents and CopilotKit in the CopilotKit documentation — end-to-end Next.js, dev server, and Deep Agent path
• CopilotKit: LangGraph features — generative UI, HITL, shared state
• LangGraph deployment — production and dev server

Best practices

• Keep the custom endpoint thin: use it to adapt CopilotKit to your graph deployment, not to duplicate business logic already inside the graph
• Send the schema explicitly: useAgentContext should describe the UI contract every time the page mounts
• Register a constrained component set: expose only the components and props you actually want the model to use
• Treat rendering as a parsing step: parse assistant content against your schema before rendering it
• Keep user messages plain: only assistant messages need the structured renderer; user messages can stay normal chat bubbles