CopilotKit

CopilotKit is a frontend and middleware layer, not an agent framework. A request goes from a framework-native frontend SDK to the Copilot Runtime, a handler you mount in your own application server; the runtime handles authentication and tool calls and opens an AG-UI session with the configured agent, which streams text, tool calls and state events back through the runtime to the UI. When the agent calls a frontend tool, the runtime relays it to a browser handler and returns the result to the agent.

The runtime holds a map of named agents, either CopilotKit's BuiltInAgent or remote AG-UI endpoints (HttpAgent) and framework adapters, and the frontend picks one by that key. Multi-agent coordination happens inside whichever framework you connect; CopilotKit renders the result. For A2A, the documented pattern wraps an AG-UI orchestrator agent in A2AMiddlewareAgent, which injects a tool the orchestrator uses to message the listed A2A agents.

Shared state lives in the agent and is synchronised to the UI through AG-UI state events and useAgent. In the open-source packages, conversation persistence is yours to run (for example LangGraph checkpointers or ADK sessions). The paid CopilotKit Intelligence service, cloud-hosted or self-hosted on Kubernetes or AWS ECS, adds AG-UI Streams (saved threads and resumable runs), user memories, automatic learning, analytics and Slack/Teams channels.

CopilotKit documents two ways to pause an agent for a person. With useHumanInTheLoop, the frontend registers a tool with a schema and a render function; when the model calls it, the chat shows your component (for example an approve button or a picker) and respond sends the user's answer back as the tool result, after which the agent continues. With useInterrupt, a LangGraph agent calls interrupt(...) in a node and the frontend renders a component for that checkpoint before the graph resumes. The docs also recommend pairing tool renderers with approval for irreversible MCP actions, and the Channels product carries approvals into Slack and Microsoft Teams.

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

// app/api/copilotkit/[[...slug]]/route.ts  (Next.js; also: npm install @ag-ui/client)
import { CopilotRuntime, createCopilotRuntimeHandler } from "@copilotkit/runtime/v2";
import { HttpAgent } from "@ag-ui/client";

// Two AG-UI agent servers you run, e.g. LangGraph and Google ADK.
// The map keys are the only names the frontend can ask for.
const runtime = new CopilotRuntime({
  agents: {
    researcher: new HttpAgent({ url: "http://localhost:8000/" }),
    writer: new HttpAgent({ url: "http://localhost:8001/" }),
  },
});
const handler = createCopilotRuntimeHandler({ runtime, basePath: "/api/copilotkit" });
export const GET = handler;
export const POST = handler;

// Client side: wrap the app in <CopilotKitProvider runtimeUrl="/api/copilotkit">
// and render <CopilotChat agentId="researcher" /> from "@copilotkit/react-core/v2".

• Needs Node.js 20 or later. The v2 APIs live under @copilotkit/react-core/v2 and @copilotkit/runtime/v2; import the stylesheet from @copilotkit/react-core/v2/styles.css.
• BuiltInAgent is CopilotKit's own agent: registering it as default replaces, rather than connects, an agent you already have. Use your framework's quickstart for runtime wiring.
• The frontend addresses agents only by the keys of the runtime's agents map, not by the agent's own name.
• A relative runtimeUrl works only when the same server hosts the app and the runtime; client-only frontends need a standalone runtime and an absolute URL.
• The runtime pins @ag-ui/mcp-apps-middleware 0.0.3, so includeTools / excludeTools raise a configuration error.
• npx copilotkit@latest create scaffolds a new project; it does not modify an existing app.

CopilotKit's team created the AG-UI protocol, and CopilotKit uses it as the wire format between its runtime and agent backends. AG-UI is the open protocol; CopilotKit is one frontend framework that speaks it.

Any backend with an AG-UI endpoint. The docs include integrations for LangGraph (Python, TypeScript, FastAPI), Google ADK, Mastra, CrewAI, Pydantic AI, Claude Agent SDK, Agno, AG2, LlamaIndex, AWS Strands, Microsoft Agent Framework and Deep Agents, plus CopilotKit's own BuiltInAgent.

MCP yes: MCP servers can be connected and MCP Apps render as UI in the chat. A2A partially: it works through an AG-UI middleware from the AG-UI repository that lets an orchestrator agent message A2A agents.

The SDKs and runtime are MIT-licensed and free. CopilotKit Intelligence has a free Developer tier and paid Pro ($39/month) and Team ($100 per seat per month) plans, with Enterprise on request.

Only as a development aid: CopilotKit publishes agent skills and a documentation MCP server that coding agents such as Claude Code and Codex can load while you build. CopilotKit does not run or drive those agents.