Skip to content
← Tools / Skills
ToolCrewAI Inc./10 trustVerified

CrewAI

Role-based multi-agent framework — crews of collaborating AI agents plus event-driven Flows, in Python.

Visit official site →

About

CrewAI models multi-agent systems as teams: agents with roles/goals/backstories, tasks, and Crews that run sequentially, hierarchically, or in parallel, plus Flows for precise event-driven control.

The official README ships first-party skills for Claude Code, Cursor, Codex, and Windsurf (verified in-repo). A Sept 2026 independent review found it the fastest route to a working multi-agent prototype. Note: GitHub shows MIT while PyPI metadata lists Apache-2.0 (per a Sept 2026 review) — a license discrepancy worth flagging.

What it can do

  • Role-based agents (role, goal, backstory) with task delegation and collaboration
  • Sequential, hierarchical, and parallel crew processes; event-driven Flows for precise control
  • Short-term, long-term, and entity memory; input/output guardrails
  • Built-in + custom tools, MCP support, knowledge sources
  • CLI scaffolding, YAML/JSONC crew definitions
  • CrewAI AMP: commercial control plane with tracing, RBAC, audit, and cost accounting

Setup

  1. Install Python 3.10–3.14.
  2. Install uv (Astral's package manager).
  3. uv tool install crewai (the CrewAI CLI).
  4. crewai create crew <project_name> — scaffolds agents/tasks/crew config.
  5. Set your model API key in the generated .env.
  6. crewai run to execute the crew.

Prerequisites

Python ≥3.10 and <3.14; uv; LLM API key.

Use cases

automationcodedatachatsearch

Works with

Claude CodeCursorCodexWindsurfGeneric API

Trust score — 9/10

/10 trustVerified

59.6k stars, very active, excellent docs, and the only entry here with verified first-party integrations for Claude Code, Cursor, Codex, and Windsurf. Docked for: (1) the MIT-vs-Apache-2.0 metadata discrepancy between GitHub and PyPI; (2) the README's telemetry section whose opt-out behavior was not independently verified; (3) growing commercial AMP upsell pressure. Vendor claims (Fortune 500 usage, agent runs) are marketing — provider-claimed.

★ 59,613 GitHub stars

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
9.6/10
Openness · license + docs
9.3/10
Manual review · sanity read + red-flag check
8.0/10
  • · 59,613 GitHub stars (log-scaled vs catalog peers)
  • · blended with ~2,442,764 monthly package downloads
  • · last push 0d ago; 100 releases/12mo; 657 open issues
  • · manual review: telemetry + upsell in quickstart; license metadata discrepancy

Sources

Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.