CopilotKit: an MIT frontend layer that streams your agent over AG-UI into React, Angular, Vue, Slack and Teams

Your agent reasons, CopilotKit renders, and AG-UI carries the events

CopilotKit does not replace the agent you built. The agent handles reasoning and execution; CopilotKit turns that work into a product experience. Events stream over AG-UI, the Agent-User Interaction protocol made by the same team, so the agent can show its work as it happens, read what the user is looking at, call tools that render UI, and pause for a person to approve the next step. The surfaces are React, Angular, Vue, React Native, Slack and Microsoft Teams, and the agent logic stays the same on each. CopilotKit now speaks AG-UI 1.0 and the A2UI 1.0 spec for declarative generative UI. The consequence: your agent has to emit AG-UI events, either natively or through one of the framework integrations, before any of the UI features work.

Three files wire a Next.js app to the v2 runtime

The hand-wired path for Next.js starts with two packages:

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

An OPENAI_API_KEY goes in .env.local. The server route at app/api/copilotkit/[[...slug]]/route.ts creates a CopilotRuntime whose default agent is a BuiltInAgent with model openai:gpt-5.4-mini, wraps it with createCopilotRuntimeHandler using basePath /api/copilotkit, and exports the handler for GET, POST, PATCH and DELETE. The page wraps its content in CopilotKitProvider with runtimeUrl set to /api/copilotkit, imports the v2 stylesheet and drops in a CopilotSidebar. Three values have to agree: the folder path of the route, basePath and runtimeUrl. Change one without the others and the sidebar has no backend to talk to. The example also exports four HTTP methods, so a proxy or middleware that only lets GET and POST through will break it.

Onboarding runs through your coding agent and always pulls @latest

The suggested quick start is to paste a prompt into Claude Code, Cursor, Codex or another coding agent, asking it to run this from the project root:

npx --yes copilotkit@latest onboard start

The command inspects your app and walks through the integration. A fresh project comes from npx copilotkit@latest create, which scaffolds in its own directory and leaves existing apps alone, and npx copilotkit@latest skills install gives the coding agent CopilotKit skills for building and debugging. All three use @latest. With v1.83.0 on 2026-10-28 and v1.84.0 on 2026-11-03, two developers onboarding a week apart can run different CLI versions and get different results. The onboarding route also means a coding agent is editing your project, so review its diff before you commit.

The Built-in Agent is OpenAI-shaped; other frameworks need their own quickstart

The three-file example uses the BuiltInAgent with an OpenAI model string and an OpenAI API key, so as written it ties the demo to OpenAI. If you already run LangGraph, CrewAI, Mastra or another framework, you swap the Built-in Agent for yours by following that framework's quickstart on docs.copilotkit.ai. That is where the real integration work sits for most teams. The repository reflects the spread: examples/integrations/ sits alongside examples for canvas, shadcn, Slack and Teams, and there is a separate sdk-python/ directory at the root. Plan for the integration step to be the longest part of adoption if your agent is not the Built-in Agent, and read the quickstart for your specific framework rather than adapting the Next.js snippet.

v1 and v2 APIs live side by side

The quick start imports from versioned subpaths: @copilotkit/runtime/v2 for CopilotRuntime, createCopilotRuntimeHandler and BuiltInAgent, and @copilotkit/react-core/v2 for CopilotKitProvider and CopilotSidebar, along with @copilotkit/react-core/v2/styles.css. The examples/ folder is split into v1/ and v2/, and the root carries codemods/, migrations.json and VERSIONING.md. In practice this means a tutorial or blog post written against the older imports can still compile while using a different API from the one the current quick start shows. When you follow outside material, check which subpath it imports from, keep a project on one API generation, and look in codemods/ and VERSIONING.md before moving an existing app across.

Channels renders one JSX tree as Block Kit in Slack and Adaptive Cards in Teams

Channels lets you write Channels JSX once and render native Block Kit in Slack and Adaptive Cards in Microsoft Teams. It ships on its own release track: channels/v0.12.0 came out on 2026-10-28, the same day as the main v1.83.0. Two version lines mean two sets of changelogs to follow, and a pre-1.0 number on the Channels side next to a 1.x number on the main packages. If you are building a Slack or Teams bot on it, pin the Channels version separately and read its release notes before each upgrade, since a 0.x package gives fewer signals about which changes are breaking. The examples/slack/ and examples/teams/ folders are the place to see the current shape of that API.

The MIT core and the production add-ons are separate offers

The repository is MIT licensed. For production, CopilotKit Intelligence adds conversations that pick up where users left off, memory across sessions, and Continual Learning, which turns real use into Skills you review and publish, learning from full-stack trajectories captured over AG-UI. CopilotKit Soar, described as AI coworkers your company owns running on your own servers and powered by OpenBot, is in early access. OpenBot itself is a separate MIT-licensed repository. These are linked to pages on copilotkit.ai rather than shipped as part of the quick start. The consequence for planning: the MIT license covers this repository, so check the terms of Intelligence and Soar on copilotkit.ai before building a production design that depends on session memory or Continual Learning.

The default build and test skip the Python, Go, Ruby and .NET runtimes

The root package.json drives an Nx workspace with pnpm. Its build and test scripts run across packages/** but exclude runtime-python, runtime-go, runtime-ruby and runtime-dotnet, along with the intelligence adapters for Python, LangGraph, ADK and the .NET agent framework. Those have their own scripts: build:native-runtimes, test:native-runtimes, build:skill-adapters and test:skill-adapters. Linting uses oxlint and formatting uses oxfmt. For contributors, a green pnpm test says nothing about the non-TypeScript runtimes; a change that touches shared protocol code has to be checked with the native-runtime and skill-adapter scripts as well, and those need the matching language toolchains installed.

Editorial conclusion

CopilotKit fits teams that already have an agent, or are happy to start with its Built-in Agent, and want chat, generative UI and human approval steps in a React, Angular, Vue, React Native, Slack or Teams front end without writing the streaming layer themselves. It is a weaker fit if you need a frozen API surface, because releases arrive about weekly and v1 and v2 imports coexist. Before committing, pin the package versions instead of relying on copilotkit@latest, decide whether you are on the v2 imports, and check the terms of Intelligence and Soar on copilotkit.ai if your production plan depends on them.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes