assistant-ui vs CopilotKit: which workflow fits?

assistant-ui vs CopilotKit: start from the interaction

For a React application whose immediate need is a composed chat interface around an existing backend, shortlist assistant-ui. For an application where an agent and users interact through shared app state, rendered tools, and approval steps, shortlist CopilotKit. Both can support agent-facing experiences; these starting points do not make the products mutually exclusive.

The assistant-ui repository describes a React chat library and backend adapters. CopilotKit's official introduction describes an app-to-agent framework. CopilotKit here is distinct from GitHub Copilot and Microsoft Copilot Studio.

Assign responsibilities before comparing features

Layer	assistant-ui	CopilotKit
UI	Compose chat primitives and message/tool displays	Build chat and interactive agent UI
State and runtime	A runtime connects UI state to the backend	Agent and app interactions can share state
Backend and protocol	Choose an adapter for your backend shape	Check the agent backend and AG-UI integration path
Persistence	Own storage or optional Assistant Cloud	Check the selected backend and optional Intelligence offering
Approval	Design the workflow and backend enforcement	Human-in-the-loop interaction is documented

Trial checks: Which components and states your users need; Which side owns a run and handles conflicts; Exact event schema, package versions, and supported events; Who owns threads, history, auth, and retention; A visible confirmation must gate the actual server action.

See assistant-ui architecture for its layer boundaries and CopilotKit's official repository for shared state and human-in-the-loop capabilities. An advertised protocol adapter is not proof that every event or tool workflow interoperates in your application.

Fit and limits

assistant-ui is a useful trial candidate when chat presentation and backend adaptation are the main work. Rendering a chat component alone does not supply your model service, access rules, or durable storage. Choose those responsibilities explicitly.

CopilotKit is a useful trial candidate when UI updates and human decisions are part of the agent workflow. Tool execution, authorization, and the semantics of shared state still need application rules. A supported backend label does not prove that your custom state or recovery path works unchanged.

Both official repositories identify their source-code license as MIT. Check the license of the exact packages and dependencies you ship. Hosted Assistant Cloud and CopilotKit Intelligence have their own service terms and plan boundaries; a library license does not make hosted services free or satisfy your data requirements.

Test one complete interaction

1. Pin package versions and select one backend. Document UI, protocol, persistence, and authorization owners.
2. Send a prompt, stream a response, render a tool result, and reload the thread. Verify retained messages and any app state.
3. Pause an action for approval. Cancel once and approve once; confirm the backend executes only the permitted action.
4. Interrupt the stream, retry after a disconnect, and switch users. Look for duplicate effects, lost history, and unauthorized state exposure.
5. Record missing events and adapter work before choosing. If considering a mixed stack, validate that combination separately.

WWW Agents has not integrated or benchmarked these products here. Performance, package compatibility, and security suitability remain unverified. Use the assistant-ui Product page for its introduction and CopilotKit docs for its official entry.