CrewAI

Overview

CrewAI is a lean, code-first Python framework for orchestrating autonomous AI agents, with no visual canvas and no dependency on LangChain. It gives developers two complementary primitives: Crews, role-based teams of agents that collaborate autonomously, and Flows, event-driven workflows with explicit state and branching for precise control.

How it works

• Define agents in agents/*.jsonc (role, goal, backstory, tools, LLM, memory) and tasks/process in crew.jsonc; crewai run loads the definition directly.
• Use Crews when you want agents to delegate and collaborate with flexible, autonomous decision-making.
• Use Flows (@start, @listen, @router decorators, plus or_/and_ conditions) when you need deterministic, event-driven control over execution paths, and call Crews as steps inside a Flow.
• Ships official skills for Claude Code, Cursor, Codex and Windsurf (crewAIInc/skills) that teach the coding agent CrewAI's own patterns for scaffolding projects and designing agents/tasks.
• Supports tools, memory, checkpointing, async execution and MCP/A2A for production-grade agents; anonymous telemetry only (no prompts/outputs collected unless you opt in).

Examples

• A hierarchical Crew of a market analyst and a researcher agent, combined in a Flow that routes to different strategies based on a confidence score.
• Automating job-posting or trip-planning pipelines from the official crewAI-examples repo.
• Installing the CrewAI skill pack so Claude Code or Cursor scaffolds a new Crew/Flow project following CrewAI's own conventions.

Installation

You need Python 3.10 or newer. The commands below use uv, a fast Python package and tool manager; install it first if you don't already have it (pip install uv, or the standalone installer from Astral's docs).

1. Install CrewAI as a global tool:

   uv tool install crewai
   

2. Scaffold a new project:

   crewai create crew <project_name>
   cd <project_name>
   

3. Install the project's own dependencies:

   crewai install
   

4. Open the generated .env file and add the API key for whichever LLM you're using (OpenAI, Anthropic, or another provider).
5. Run the crew:

   crewai run
   

If you're adding CrewAI to an existing coding agent instead of running it standalone, install the official skill pack for Claude Code, Cursor, Codex or Windsurf from the crewAIInc/skills repo so the agent follows CrewAI's own project conventions.