CopilotKit is an open-source full-stack SDK for building agent-native applications with generative UI, shared state, and human-in-the-loop workflows. With 31,000+ GitHub stars, 4,000+ forks, and a TypeScript-first design for React and Angular, CopilotKit has emerged as the dominant frontend layer for embedding AI agents into web applications. The team behind the project also created the AG-UI Protocol, which has been adopted by Google, LangChain, AWS, Microsoft, Mastra, and PydanticAI as a standard for agent-to-user-interface communication. ## Why CopilotKit Matters Most agent frameworks today focus on backend orchestration: LangGraph, CrewAI, Mastra, and similar projects describe how agents reason, plan, and call tools. CopilotKit fills a complementary and often-overlooked gap on the frontend. It defines how those agents talk to a user interface, render live components, share state with React, and pause to ask for human input. By pairing this frontend stack with the AG-UI Protocol, CopilotKit makes it practical to ship production agent applications that feel like first-class products rather than chat boxes bolted onto a backend. ## Generative UI and Backend Tool Rendering The core capability is generative UI: agents can dynamically render and update React or Angular components at runtime based on user intent and agent state. Three patterns are supported - static rendering via the AG-UI Protocol, declarative rendering via A2UI, and open-ended rendering via MCP Apps and Open JSON. Backend tool rendering lets agents invoke server-side tools whose results come back as UI components rendered directly in the client, eliminating the need for custom plumbing on each tool call. ## Shared State and the useAgent Hook CopilotKit introduces a synchronized state layer that both agents and UI components read from and write to in real time. The useAgent hook (a proper superset of the older useCoAgent) sits directly on AG-UI and gives developers fine-grained control over the agent connection, including programmatic access to agent state and the ability to mutate state from the UI. This makes it natural to build interfaces where users and agents collaborate on the same documents, dashboards, or workflows. ## Human-in-the-Loop Workflows Agents can pause execution to request user input, confirmation, or edits before continuing. This is essential for any production scenario where the agent's action has real consequences - submitting a form, sending an email, modifying a database, or executing a transaction. CopilotKit's approval and clarification primitives are integrated into the protocol layer rather than bolted on as application code. ## AG-UI Protocol Ecosystem The AG-UI Protocol that CopilotKit authored is now an industry-standard agent-to-UI interaction layer. First-party integrations exist for LangGraph, CrewAI, Mastra, PydanticAI, and many other agent frameworks. This means a CopilotKit application can swap backend frameworks without rewriting the UI, and agents written against AG-UI can be embedded into any AG-UI-aware frontend. ## Developer Experience Getting started is a single command: npx copilotkit@latest create for new projects, or npx copilotkit@latest init for existing ones. The CLI installs core packages, configures providers, wires agents to the UI, and produces a deployment-ready app. The same package supports both React and Angular, and the framework integrates cleanly with Next.js, making it well suited for modern full-stack TypeScript projects. ## Limitations CopilotKit is opinionated about React and Angular, so teams using Vue, Svelte, or non-component frontend stacks will need to wait for community ports or adopt AG-UI directly. Generative UI quality depends heavily on the underlying agent and model, and complex stateful UIs require careful component design to avoid stale state. The MIT-licensed core is free, but the team also offers a managed Copilot Cloud and an Enterprise Intelligence Platform for teams that prefer hosted infrastructure.