getting-started

Use LLM.call() for single prompts without tools, Agent.kickoff() for one agent with tools, Crew.kickoff() for multi-agent collaboration, and Flow when you need state management, routing, or multiple crews. For production apps, always start with a Flow since it can embed all other abstractions.

Run 'crewai create flow <project_name>' with underscores in the name, never hyphens. The CLI generates correct imports, directory structure, and pyproject.toml. Then edit the generated agents.yaml, tasks.yaml, crew.py, and main.py rather than writing files from scratch.

The variable was not passed to kickoff(). CrewAI uses single-brace {variable} placeholders in YAML that are replaced at runtime via crew.kickoff(inputs={"topic": "..."}). Missing inputs leave the literal placeholder in the prompt.

Use a conversational Flow with ConversationState and call flow.handle_turn(message, session_id=...) for each user line. Route turns with route_turn() plus @listen("ROUTE") handlers, call append_assistant_message(reply) in handlers, and finalize traces with finalize_session_traces().

Prefer official MCP servers when one exists for the service, since providers maintain them and they stay current with API changes. Attach them via the agent's mcps field. Fall back to native crewai_tools like SerperDevTool only when no official MCP server is available.

The project name likely contains hyphens, which are not valid Python identifiers. Recreate the project with underscores, for example 'crewai create flow my_project', and verify import paths match the generated package structure.