copilotkit-debug

First verify the agent is registered by curling the /info endpoint and comparing the agent name with your agentId prop. Then inspect the SSE stream in the browser Network tab for RunErrorEvent or a stalled stream, and check server logs for missing LLM API keys or invalid model strings.

CopilotKit CORS errors occur when the runtime origin config does not match your frontend. By default all origins are allowed without credentials; if you need cookies, set an explicit origin with credentials: true in createCopilotEndpoint and pass credentials="include" to CopilotKitProvider.

Premature stream termination is usually caused by hosting platform timeouts (Vercel defaults to 30s), an uncaught agent exception, or client disconnection. Check for a RunErrorEvent before the cutoff and consider Intelligence mode for long-running agents.

AGENT_NOT_FOUND means the agentId passed to CopilotChat or useAgent does not match any key in the runtime's agents map. Agent names are case-sensitive, so query the /info endpoint to list registered agents and align the names exactly.

Frontend tools fail when the tool name does not exactly match what the agent calls, the useFrontendTool hook is not mounted before the run starts, or the execute handler throws. Check the SSE stream for ToolCallStartEvent and look for tool_not_found or tool_handler_failed error codes.

Yes, via LangGraphAgent, but there is a known issue (#3519) where the Python SDK dispatches events with a copilotkit_ prefix that ag-ui-langgraph does not recognize, causing emitted state and messages to be silently dropped. Updating ag-ui-langgraph or patching the event name mapping resolves it.