CopilotKit Setup

Prerequisites

Live Documentation (MCP)

This plugin includes an MCP server (copilotkit-docs) that provides search-docs and search-code tools for querying live CopilotKit documentation and source code.

• Claude Code: Auto-configured by the plugin's .mcp.json -- no setup needed.
• Codex: Requires manual configuration. See the copilotkit-debug skill for setup instructions.

Environment

Before starting setup, verify:

1. Node.js >= 18 (required for fetch globals used by the runtime)
2. An AI provider API key (one of: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY)
3. A React-based frontend (Next.js App Router, Next.js Pages Router, Vite + React, or Angular)
4. A backend capable of running the runtime (same Next.js app via API routes, or a standalone Express/Hono server)

Framework Detection

Before generating any code, detect the project's framework by checking files in the project root. See references/framework-detection.md for the full decision tree.

Quick summary:

Setup Workflow

Step 1: Install packages

All packages use the @copilotkit namespace. The v2 API lives as subpath exports on the published packages.

Frontend + backend in the same Next.js app:

npm install @copilotkit/react-core @copilotkit/runtime hono

Frontend only:

npm install @copilotkit/react-core

Backend runtime only:

npm install @copilotkit/runtime hono

For standalone Express backends, install Express adapter dependencies instead of hono:

npm install @copilotkit/runtime express dotenv zod
npm install -D @types/express tsx typescript

(createCopilotExpressHandler enables CORS internally, so you do not need to install cors yourself. dotenv and zod are used by the example asset.)

Step 2: Choose a runtime mode, then configure the runtime

The runtime is the server-side component that manages agent execution. See references/runtime-architecture.md for details.

Decide the mode before writing any runtime code. The mode changes how the runtime is constructed, so retrofitting it later means rewriting this file.

Ask the user which they want, defaulting to managed Intelligence. State the prerequisites plainly so the choice is informed:

Managed Intelligence requires a CopilotKit account (free; npx copilotkit login opens the browser) and a project API key. In exchange, threads are durable across restarts and deploys, you get the dashboard, and Slack/Teams Channels become available -- Channels are not available in SSE mode at all.

Self-hosted SSE requires nothing beyond an AI provider key. Thread state lives in memory in the process that served the request, so it is lost on restart and is not shared across replicas. Everything in this skill works in SSE mode; it is a supported path, not a dead end.

If the user picks managed Intelligence, use the Intelligence runtime blocks below and then complete Step 6. If they pick SSE, use the SSE blocks and skip Step 6's server-side wiring.

There are two endpoint styles:

1. Multi-route (Hono) -- uses createCopilotHonoHandler. Requires a catch-all route ([[...slug]] in Next.js). Each operation (run, connect, stop, info, transcribe, threads) gets its own HTTP path.
2. Single-route (Hono or Express) -- uses createCopilotHonoHandler({ ..., mode: "single-route" }) or createCopilotExpressHandler({ ..., mode: "single-route" }). All operations go through a single POST endpoint with method multiplexing.

Next.js App Router (recommended: multi-route with Hono)

Create src/app/api/copilotkit/[[...slug]]/route.ts:

import {
  CopilotRuntime,
  createCopilotHonoHandler,
  InMemoryAgentRunner,
  BuiltInAgent,
} from "@copilotkit/runtime/v2";
import { handle } from "hono/vercel";

const agent = new BuiltInAgent({
  model: "openai/gpt-4o",
  prompt: "You are a helpful AI assistant.",
});

const runtime = new CopilotRuntime({
  agents: {
    default: agent,
  },
  runner: new InMemoryAgentRunner(),
});

const app = createCopilotHonoHandler({
  runtime,
  basePath: "/api/copilotkit",
});

export const GET = handle(app);
export const POST = handle(app);
// PATCH/DELETE are used by thread operations (useThreads); export them too
// so the multi-route handler can serve them when you enable Intelligence/threads.
export const PATCH = handle(app);
export const DELETE = handle(app);

Next.js App Router with managed Intelligence

Same file and same handler as above; only the runtime construction differs. intelligence is what selects Intelligence mode, and identifyUser is required with it.

import {
  CopilotRuntime,
  CopilotKitIntelligence,
  createCopilotHonoHandler,
  BuiltInAgent,
} from "@copilotkit/runtime/v2";
import { handle } from "hono/vercel";

const agent = new BuiltInAgent({
  model: "openai/gpt-4o",
  prompt: "You are a helpful AI assistant.",
});

const intelligence = new CopilotKitIntelligence({
  // Server-side secret. `apiUrl`/`wsUrl` default to CopilotKit's managed
  // platform, so most projects set only the key.
  apiKey: process.env.COPILOTKIT_API_KEY!,
});

const runtime = new CopilotRuntime({
  agents: { default: agent },
  intelligence,
  // REQUIRED in Intelligence mode: it decides whose threads these are. Resolve a
  // real authenticated user from the request -- with a hardcoded id, every visitor
  // shares one thread history.
  identifyUser: (request) => resolveUserFromSession(request),
});

const app = createCopilotHonoHandler({
  runtime,
  basePath: "/api/copilotkit",
});

export const GET = handle(app);
export const POST = handle(app);
export const PATCH = handle(app);
export const DELETE = handle(app);

Notes that matter:

• Do not pass runner. Intelligence mode supplies IntelligenceAgentRunner itself; passing InMemoryAgentRunner is what keeps threads in memory.
• identifyUser is not optional. It is the thread-ownership boundary. A stub like () => ({ id: "demo", name: "Demo" }) is fine for a local spike and wrong for anything multi-user.
• apiUrl and wsUrl are separate hosts. They default to https://api.intelligence.copilotkit.ai and wss://realtime.intelligence.copilotkit.ai. The realtime plane cannot be derived from the API plane by swapping the scheme, so override both or neither -- overriding one points the two planes at different deployments.

Next.js App Router (alternative: single-route)

Create src/app/api/copilotkit/route.ts:

import {
  CopilotRuntime,
  createCopilotHonoHandler,
  InMemoryAgentRunner,
  BuiltInAgent,
} from "@copilotkit/runtime/v2";
import { handle } from "hono/vercel";

const agent = new BuiltInAgent({
  model: "openai/gpt-4o",
  prompt: "You are a helpful AI assistant.",
});

const runtime = new CopilotRuntime({
  agents: {
    default: agent,
  },
  runner: new InMemoryAgentRunner(),
});

const app = createCopilotHonoHandler({
  runtime,
  basePath: "/api/copilotkit",
  mode: "single-route",
});

export const POST = handle(app);

When using single-route, the frontend must set useSingleEndpoint on the provider (see Step 3).

Standalone Express Server

Create src/index.ts:

import express from "express";
import { CopilotRuntime, BuiltInAgent } from "@copilotkit/runtime/v2";
import { createCopilotExpressHandler } from "@copilotkit/runtime/v2/express";

const agent = new BuiltInAgent({
  model: "openai/gpt-4o",
});

const runtime = new CopilotRuntime({
  agents: {
    default: agent,
  },
});

const app = express();

app.use(
  "/api/copilotkit",
  createCopilotExpressHandler({
    runtime,
    basePath: "/",
    mode: "single-route",
  }),
);

const port = Number(process.env.PORT ?? 4000);
app.listen(port, () => {
  console.log(
    `CopilotKit runtime listening at http://localhost:${port}/api/copilotkit`,
  );
});

For multi-route Express, omit the mode option (multi-route is the default) -- createCopilotExpressHandler is the same factory for both styles (imported from @copilotkit/runtime/v2/express).

Standalone Hono Server (non-Vercel)

import {
  CopilotRuntime,
  createCopilotHonoHandler,
  BuiltInAgent,
} from "@copilotkit/runtime/v2";
import { serve } from "@hono/node-server";

const runtime = new CopilotRuntime({
  agents: {
    default: new BuiltInAgent({ model: "openai/gpt-4o" }),
  },
});

const app = createCopilotHonoHandler({
  runtime,
  basePath: "/api/copilotkit",
});

serve({ fetch: app.fetch, port: 8787 });

Requires @hono/node-server:

npm install hono @hono/node-server

Step 3: Set up the frontend provider

Wrap your application with CopilotKit from @copilotkit/react-core/v2.

Which provider component? Always use CopilotKit imported from @copilotkit/react-core/v2. It is the compatibility bridge across v1 and v2 and a strict superset of the other provider APIs. Do not use CopilotKit from the package root (@copilotkit/react-core, legacy v1) or CopilotKitProvider from /v2 (a subset of the functionality).

Important: Import the stylesheet in your root layout:

import "@copilotkit/react-core/v2/styles.css";

Next.js App Router

In src/app/page.tsx (or a client component):

"use client";

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

export default function Home() {
  return (
    // useSingleEndpoint={false} matches the multi-route backend above.
    // The v1-compat CopilotKit bridge defaults useSingleEndpoint to true,
    // which would 404 against multi-route endpoints.
    <CopilotKit runtimeUrl="/api/copilotkit" useSingleEndpoint={false}>
      <div style={{ height: "100vh" }}>
        <CopilotChat />
      </div>
    </CopilotKit>
  );
}

Connecting to an external runtime

When the runtime runs on a separate server (e.g., Express on port 4000):

<CopilotKit runtimeUrl="http://localhost:4000/api/copilotkit" useSingleEndpoint>
  {children}
</CopilotKit>

Set useSingleEndpoint when the backend uses single-route endpoints (createCopilotHonoHandler or createCopilotExpressHandler with mode: "single-route").

CopilotKit key props

Step 4: Add a chat UI component

CopilotKit provides three pre-built chat layouts (all imported from @copilotkit/react-core/v2):

Example with sidebar:

import { CopilotKit, CopilotSidebar } from "@copilotkit/react-core/v2";

<CopilotKit runtimeUrl="/api/copilotkit" useSingleEndpoint={false}>
  <YourApp />
  <CopilotSidebar
    defaultOpen
    width="420px"
    labels={{
      modalHeaderTitle: "AI Assistant",
      chatInputPlaceholder: "Ask me anything...",
    }}
  />
</CopilotKit>;

Step 5: Set environment variables

Provider API keys are secrets. Store them in environment variables -- never hardcode them in source or commit them to version control. Create a .env.local (Next.js) or .env file:

OPENAI_API_KEY=<your-openai-api-key>

Make sure your .gitignore excludes env files (.env, .env.local, .env*.local) so keys are never committed. In production, supply keys through your platform's secret manager (Vercel/Netlify environment variables, AWS Secrets Manager, etc.) rather than a checked-in file.

The BuiltInAgent automatically resolves API keys from these environment variables based on the model prefix:

• openai/* models read OPENAI_API_KEY
• anthropic/* models read ANTHROPIC_API_KEY
• google/* models read GOOGLE_API_KEY

If you need to pass apiKey explicitly, always source it from the environment (apiKey: process.env.OPENAI_API_KEY) -- never inline a literal key.

Step 6: Connect to CopilotKit Intelligence

Skip this step only if the user chose self-hosted SSE in Step 2.

Intelligence has two halves and they use different credentials. Getting them mixed up is the most common setup mistake here.

1. Sign in and create a project.

   bashCopy

   npx copilotkit login
   npx copilotkit project select
   

   login opens the browser and stores a local CLI session. project select picks or creates a hosted project and records it in .copilotkit/project.json. Verify the available commands with npx copilotkit --help if a version differs.

2. Set the server-side project API key. project select provisions one; you can also copy it from the dashboard.

   Copy

   # .env.local (Next.js) or .env
   COPILOTKIT_API_KEY=cpk-...
   

   This is a secret. It has no NEXT_PUBLIC_/VITE_ prefix on purpose -- prefixing it would ship it to the browser. It is read by the CopilotKitIntelligence client you wired in Step 2.

   CLI-scaffolded projects use INTELLIGENCE_API_KEY for the same value; either name is fine as long as the runtime reads the one you set.

3. Set the public license key and pass it to the provider. Unlike the API key, this one is a public project identifier and is meant to reach the client:

   Copy

   NEXT_PUBLIC_COPILOTKIT_LICENSE_KEY=<your-license-key>
   

   tsxCopy

   <CopilotKit
     runtimeUrl="/api/copilotkit"
     publicLicenseKey={process.env.NEXT_PUBLIC_COPILOTKIT_LICENSE_KEY}
   >
   

   The NEXT_PUBLIC_/VITE_ prefix is required because the key is read on the client.

4. Confirm durable threads actually work. Send a message, restart the dev server, and reload. The thread should still be there. If it is not, the runtime is still in SSE mode -- check that intelligence is passed and that no runner overrides it.

See references/telemetry-setup.md for what the license key enables and how to opt out.

Connecting Slack or Microsoft Teams

Channels let an agent answer in Slack or Teams. They require the Intelligence runtime -- channels is not available in SSE mode -- and a long-running host, because activation opens a persistent connection.

Use the copilotkit-channels skill for this. It covers declaring a Channel, the long-running host requirement, and which mounts start activation on their own versus which wait for an explicit channels.ready() call. It builds on the wiring from this step.

For a new managed Teams app, create the Channel draft in Intelligence first and use either the recommended browser-issued Fast CLI command (channels add --project-id … --channel-id … --adapter teams --provision) or the peer Guided manual path. Do not scaffold Azure Bot resources or place Microsoft credentials, custom icons, or generated packages in the project. The provider's Created and installed result is separate from this code half; only a running host plus a real Teams interaction verifies the integration end to end.

Step 7: Verify the setup

1. Start the dev server
2. Open the app in a browser
3. The chat UI should render and connect to the runtime
4. Send a test message -- you should receive an AI response
5. Check the runtime's info endpoint to confirm it reports available agents. For multi-route handlers this is GET /api/copilotkit/info; for single-route handlers (mode: "single-route", e.g. the Express example) it is a POST to the base path with body { "method": "info" } (a plain GET will not return agent info — the Hono single-route handler answers 405, and the Express single-route router has no GET route so it falls through to a 404)

Security notes

Keep these in mind as you wire up a real deployment:

• Secrets stay server-side and in env vars. Provider API keys (OPENAI_API_KEY, etc.) are read by the runtime/agent on the server. Never expose them to the browser, hardcode them, or commit them -- store them in environment variables or a secret manager (see Step 5). The CopilotKit license key is the one client-side value, and it is a public project identifier, not a secret.
• Treat all chat input as untrusted. Chat messages flow from the frontend through the CopilotRuntime endpoint into the agent's LLM context. They are user-controlled and can attempt prompt injection -- including indirect injection via content the agent fetches (web pages, documents, tool results). Do not assume the model will only do what your system prompt intends.
• Give server-side tools least privilege. A defineTool's execute function runs with your server's authority. Validate every argument (the zod parameters schema is your first gate), scope each tool to the narrowest action it needs, and enforce your own authorization inside the execute function for anything sensitive (database writes, payments, file access) rather than trusting that the model called it correctly.
• Authenticate the runtime endpoint. The runtime route is a public HTTP endpoint by default. Put your app's auth in front of it so only authorized users can drive the agent and consume provider credits.

Quick Reference

Package map

Endpoint factory functions

The createCopilotEndpoint, createCopilotEndpointSingleRoute, createCopilotEndpointExpress, and createCopilotEndpointSingleRouteExpress names are deprecated aliases of the two factories above. Prefer the handler factories with the mode option.

Runtime classes

Channels require the Intelligence runtime and a long-running host. See the copilotkit-channels skill.

Agent runners

Supported models (BuiltInAgent)

Format: "provider/model-name" string or a Vercel AI SDK LanguageModel instance.

OpenAI: openai/gpt-5, openai/gpt-5-mini, openai/gpt-4.1, openai/gpt-4.1-mini, openai/gpt-4.1-nano, openai/gpt-4o, openai/gpt-4o-mini, openai/o3, openai/o3-mini, openai/o4-mini

Anthropic: anthropic/claude-sonnet-4.5, anthropic/claude-sonnet-4, anthropic/claude-3.7-sonnet, anthropic/claude-opus-4.1, anthropic/claude-opus-4, anthropic/claude-3.5-haiku

Google: google/gemini-2.5-pro, google/gemini-2.5-flash, google/gemini-2.5-flash-lite

Any string is accepted (for custom/unlisted models); the provider is parsed from the prefix before /.