What it is
CrewAI organises agents into Crews: each agent gets a role, a goal and tools, and the crew works through a list of tasks together. It is written in Python (3.10 to 3.13), does not depend on LangChain, and calls the OpenAI API unless you configure another model.
Flows add a more predictable layer on top: event-driven steps with state, retries and branching, where you can call a crew, a single agent or plain code. Flows are how teams take CrewAI to production, and the hosted platform runs the same model.
The company sells CrewAI AMP, a control plane for deploying and monitoring crews. Its free Basic tier allows 50 workflow executions a month; SSO, RBAC, PII redaction and deployment in your own VPC or infrastructure are Enterprise only, priced by quote.
Our review
CrewAI is the quickest route from an idea to several cooperating agents in Python, and its own product direction admits that free-form crews are not enough for production. It has about 60,000 GitHub stars and shipped 1.15.28 on 14 October 2026; the release cadence is near weekly, which is good for fixes and tiring for anyone pinning versions.
The role, goal and task model is easy to explain to a non-engineer, which is much of its appeal. The cost shows up later. Atlan's 2026 explainer says production complaints cluster around debugging opacity and token cost: a hierarchical crew fans out into many model calls, and default logging is noisy enough that most teams add structured logging before they can debug it. Flows answer this with explicit steps and state, at which point the code starts to look like LangGraph with different names.
Security needs attention. In March 2026 CERT/CC published VU#221883, four CVEs covering a sandbox escape in the Code Interpreter tool when Docker was unavailable, SSRF in the RAG search tools and unchecked file reads in the JSON loader. CrewAI's response, logged by CERT in May, was to remove the Code Interpreter tool entirely, deprecate allow_code_execution and add path and URL validation to more than 20 tools; CERT still lists all four as affected because no fixed version number was given.
The library also sends anonymous telemetry by default. The README says it excludes prompts, task text and outputs unless you set share_crew to True, and OTEL_SDK_DISABLED=true turns it off; a regulated team should set that before the first run rather than after the security review.
On price, AMP's public page now shows only a free tier of 50 executions and an Enterprise quote, so a buyer cannot cost a mid-sized deployment without a sales call, whereas LangSmith publishes per-seat and per-vCPU rates.
Where it earns its place
A two-person data team at a mid-sized e-commerce firm
Prototype a research-and-summary pipeline for weekly competitor reports
A three-agent crew (researcher, analyst, writer) runs from a YAML config and a single Python file, usually within an afternoon. The library is free; the bill is model tokens, which grow with every hand-off between agents, so cap iterations per task.
A platform engineer at a bank moving an agent pilot to production
Make an agent workflow predictable enough to pass change control
Rewrite the crew as a Flow with typed state and explicit steps, set OTEL_SDK_DISABLED=true, and keep code execution in an external sandbox such as E2B, since the built-in Code Interpreter tool was removed after the 2026 CVEs.
An operations director at a 300-person insurer
Get managed hosting, SSO and audit controls for agent workflows
The free AMP Basic tier allows 50 executions a month, enough for a demo only. SSO (Entra ID, Okta), RBAC, PII redaction and VPC deployment are Enterprise features with no published price, so budget for a negotiated contract.
Before you commit
- Setup
- uv tool install crewai, then crewai create crew <name> scaffolds a project; it needs Python 3.10 to 3.13 and an OpenAI key by default, or another provider or Ollama configured instead.
- Integrations
- Any LLM through its model configuration, a crewai-tools package of search, scraping and RAG tools, MCP servers, and tracing through OpenTelemetry-based tools such as OpenLLMetry.
- Data and privacy
- MIT licence with no commercial limits. Anonymous telemetry is on by default (disable with OTEL_SDK_DISABLED=true); share_crew=True sends task text and outputs. CERT VU#221883 covers four 2026 CVEs, which CrewAI says are fixed in current releases.
- Cost at scale
- Self-hosted, the cost is model tokens plus your own compute, and multi-agent crews multiply tokens per task. AMP Basic is free for 50 executions a month; anything beyond that is an Enterprise quote.
- Limits
- About 835 open issues and pull requests, near-weekly releases, and weaker step-level control and checkpointing than LangGraph for long-running or human-approved workflows.
What's good
- Role, goal and task model a non-engineer can follow
- Flows add explicit steps, state and retries
- No LangChain dependency
- Large example library and community
What's not
- Hard to see which prompts reach the model in big crews
- Agent-to-agent chatter raises token bills
- Anonymous telemetry on by default
- AMP pricing jumps from 50 free executions to a sales quote
Best for
Python developers who want a multi-agent workflow running quickly and can accept less low-level control.
Alternatives
CrewAI: common questions
- Is CrewAI worth it?
- Fast to a working multi-agent demo and harder to steer the further it travels from one, so production teams end up writing Flows rather than Crews. We rate it 4.1 out of 5.
- How much does CrewAI cost?
- Free and open source under the MIT licence. CrewAI AMP: Basic free with 50 workflow executions a month; Enterprise (more executions, SSO, RBAC, VPC or on-prem deployment) by quote.
- What are the downsides of CrewAI?
- Hard to see which prompts reach the model in big crews; agent-to-agent chatter raises token bills; anonymous telemetry on by default; aMP pricing jumps from 50 free executions to a sales quote.
- Who is CrewAI best for?
- Python developers who want a multi-agent workflow running quickly and can accept less low-level control.
- How much does CrewAI cost for a team or in production?
- Self-hosted, the cost is model tokens plus your own compute, and multi-agent crews multiply tokens per task. AMP Basic is free for 50 executions a month; anything beyond that is an Enterprise quote.
- Is CrewAI safe for company data?
- MIT licence with no commercial limits. Anonymous telemetry is on by default (disable with OTEL_SDK_DISABLED=true); share_crew=True sends task text and outputs. CERT VU#221883 covers four 2026 CVEs, which CrewAI says are fixed in current releases.
- What are the best CrewAI alternatives?
- The closest alternatives we have reviewed are LangGraph, Agno, Microsoft Agent Framework and Google Agent Development Kit (ADK).
Answers checked Oct 23, 2026.
Make CrewAI? Show this rating on your site
