CrewAI
CrewAI is an enterprise multi-agent platform that lets business teams build AI agent workflows in a visual editor while engineers work code-first in Python, with a control plane that traces, governs and audits every run.
What is CrewAI?
CrewAI exists in two connected forms: an open source Python framework for building crews of AI agents, and AMP, the Agent Management Platform that runs, governs and improves those agents in production. The company frames it as an enterprise build-and-runtime layer, designed to close the gap between business teams who know which processes deserve automating and engineers who can actually ship them.
The building block is an agent described by a role, a goal and a backstory, optionally given tools, memory, reasoning and checkpointing. Several agents form a crew. Crews run through one of two processes: sequential, where each task feeds its output into the next, or hierarchical, where an automatically created manager agent delegates work and validates results — a mode that requires a configured manager LLM. Memory comes in four kinds (short-term, long-term, entity and contextual) and is presented as a cognitive layer that resolves contradictions and forgets deliberately. Checkpointing captures runtime state at every step, so a workflow can be replayed from a chosen point or forked with different inputs. The framework speaks MCP, supports custom MCP servers and the agent-to-agent protocol, and ships native E2B and Daytona sandboxes for code execution.
Commercially, the product is organised around four pillars. Discovery cross-references a customer's tickets, chats, applications and workflows against patterns drawn from billions of agent runs, then ranks automation opportunities by effort, value and readiness. Build covers Crew Studio, a no-code editor with an AI copilot whose output exports to Python, alongside the code-first API. Govern is the Control Plane, which sits in the execution path of every workflow to deliver real-time tracing with full cost accounting, role-based access control, immutable audit trails, human approval gates and runtime PII redaction. Optimize turns production runs into training data, adds multi-LLM testing with runtime model swapping, and scores outputs for hallucination.
Twenty-eight integrations are documented, spanning Google Workspace, Microsoft 365, Salesforce, HubSpot, Slack, Jira, Zendesk, Snowflake, Databricks, Stripe and Shopify, with triggers that start crews from incoming events. Deployment runs on CrewAI cloud, a dedicated VPC or the customer's own infrastructure.
What it does
- Identify which processes are worth automating by matching your tickets, chats, apps and workflows against patterns observed across billions of agent runs
- Build multi-agent workflows either from a prompt-and-canvas visual editor or code-first in Python, with role-based agents, memory and checkpointing
- Run agents in production on CrewAI cloud, a dedicated VPC or your own infrastructure
- Trace every LLM call, tool call and memory read in real time, with full cost and token accounting
- Enforce governance at runtime through role-based access control, immutable audit logs, enterprise IAM, PII redaction and human approval gates
- Connect agents to business systems through twenty-eight documented integrations and event-based triggers
- Test several LLMs against the same workflow, swap the model at runtime, and score outputs for hallucination
When to use CrewAI / When not to
A quick filter to help you decide if CrewAI is the right fit.
When to use CrewAI
- Software and machine learning engineers who want a Python framework for orchestrating crews of role-based agents rather than wiring one up themselves
- Platform and AI teams whose problem is no longer building agents but governing the ones already running, with role-based access, immutable audit trails and PII redaction
- Operations analysts and business process owners who need to automate long-horizon work through a visual canvas without writing code
- Large regulated organisations that require SSO, deployment inside their own VPC or infrastructure, and controls suited to mission-critical environments
- Support, sales, marketing, finance and HR teams looking to automate ticket handling, lead enrichment or document-heavy internal processes
When not to use CrewAI
- Individuals or small teams looking for a mobile app: there is no iOS or Android client at all
- Buyers who need a published price before talking to anyone, since everything above the free tier is quoted as Custom
- Teams intending to run production volume on the free plan, which is capped at two automations and fifty executions per month with no overage
- Organisations that need documented GDPR commitments, a data processing agreement or a named hosting region before signing anything
- Non-English-speaking users, as the interface and the entire documentation set are published in English only
How to use CrewAI
A typical end-to-end flow, from setup to results.
- Create an account at app.crewai.com, or install the open source framework locally with uv if you prefer to start from code
- For the code route, configure your model API keys and set up the CrewAI CLI for local development
- Define your agents in YAML, in Python or both, giving each a role, a goal and a backstory, then attach an LLM, tools, memory or reasoning as needed
- Define the tasks and choose the process: sequential for an ordered chain, hierarchical if you want a manager agent to delegate and validate, in which case set the manager LLM
- Alternatively, build the whole crew in Crew Studio, the visual editor with an AI copilot, and export it to Python when you want to take over in code
- Connect the integrations and triggers your workflow needs, from Gmail and Slack to Salesforce, HubSpot, Snowflake or Zapier
- Deploy through one of the three documented routes: the GitHub integration, directly from Crew Studio, or the CLI
- Call the deployed crew through its generated REST endpoints, or export it as an MCP server or a UI component
- Watch it run in the Control Plane: traces, token counts, cost, performance metrics, hallucination scores and deployment history, with OpenTelemetry export if you have your own stack
- Tighten it over time by training agents on production runs, testing several LLMs against the same workflow, and scheduling recurring executions with cron
Pros & Cons
Pros
- One product serves both audiences: the visual workflow exports to Python, so a business prototype does not have to be rebuilt from scratch by engineers
- The framework is genuinely open source and installable without an account, with a claimed community of more than 100,000 developers behind it
- A permanent free plan gives access to the visual editor, the GitHub integration and fifty workflow executions a month
- Observability is unusually detailed for this category: per-call tracing, cost, token counts, hallucination scores and OpenTelemetry export
- Governance is built into the execution path rather than bolted on, with access control, audit trails and human approval gates
- Deployment is not locked to the vendor's cloud: a dedicated VPC or the customer's own infrastructure are both supported
- Named customer references with published figures, including PwC, IBM, AWS, Docusign and Gelato, make the claims checkable
Cons
- No public pricing above the free tier: the Enterprise plan is quoted as Custom, with Request Info and Request trial as the only paths
- The free plan is tightly capped at two automations and fifty executions a month, with no overage available
- Enterprise connectors, dedicated VPC, SSO, role-based access control, dedicated support, onboarding and training are all reserved for the paid tier
- The GDPR is never mentioned anywhere on the site, and there is no data processing agreement, no subprocessor list and no Article 27 representative
- The terms explicitly allow CrewAI to use customer data to train and develop models, with no documented way to opt out
- No hosting country or region is disclosed, and the trust centre that might answer this renders nothing without JavaScript
- Contact runs through a form only: no postal address, no support email, and the interface and documentation are English-only
Pricing & Plans
A permanent free plan is available, not merely a trial, and it includes the visual editor, the AI copilot, the GitHub integration, two automations and fifty workflow executions per month. Above it, only one paid tier is offered, called Enterprise, and its price is not published: the pricing page shows Custom and directs the visitor to Request Info or Request trial. No entry-level amount or currency can therefore be reported.
- visual editor and AI copilot
- GitHub integration
- 50 workflow executions per month
- 2 automations
- private agent and tool repositories
- workflow templates
- export as MCP server and as UI component
- CrewAI cloud infrastructure
- everything in the free plan
- plus enterprise connectors
- governance from the start with SSO
- role-based access control
- workload identity
- PII redaction and policies
- deployment on CrewAI cloud
- a dedicated VPC or the customer's own infrastructure
Data, GDPR & hosting
A consolidated view of how CrewAI handles your data.
GDPR overview
There is no mention of the GDPR anywhere on this site. A full-text search across every page collected, including the archived HTML, returns zero occurrences of the regulation: no compliance claim, no lawful basis, no data subject rights section, no standard contractual clauses, no Article 27 representative and no designated data protection officer. What exists instead is a privacy policy effective 24 September 2025 offering narrower choices — marketing unsubscribe, cookie blocking, advertising opt-outs, and deletion of Google-derived data on request to privacy@crewai.com. The policy does describe security safeguards: encryption in transit and at rest, role-based access controls and continuous monitoring. A Vanta-powered trust centre exists at trust.crewai.com, but it renders nothing without JavaScript, so no certification or commitment could be verified from it.
Who owns the data?
Between the parties, the customer keeps all right, title and interest in Customer Data, which the terms define as everything uploaded to the platform plus the Outputs the agents generate. CrewAI keeps ownership of the platform itself. The customer grants CrewAI a non-exclusive, royalty-free licence, sub-licensable only to its contractors and service providers, to use that data for three purposes: operating the platform, improving and developing products and services including by training and developing models and algorithms, and creating aggregated or anonymised data, which the terms explicitly place outside the definition of Customer Data. Feedback is treated separately and far more broadly: it is licensed perpetually, irrevocably and worldwide, and is not confidential.
Reuse rights
Because ownership of Customer Data and of the Outputs stays with the customer, reusing them commercially or otherwise does not require CrewAI's permission. The terms attach conditions rather than restrictions on reuse: the customer must apply discretion and independent judgment before relying on, sharing or using any Output, and must monitor actions the agents take on its behalf. CrewAI warns that Outputs may not be unique compared with those produced for other users, may carry biases present in training data, and may be inaccurate on tasks requiring reasoning or judgment; it accepts no liability for their use. Separately, the customer must warrant it holds the rights to the data it supplies. The platform itself may not be reverse-engineered, resold as a service bureau, or used to build a competing product.
Data retention & training
Hosting summary
No hosting country or region is named anywhere: not in the privacy policy, not in the terms of use, not on the product pages. What is documented is architectural rather than geographic. The free plan runs on CrewAI cloud; the Enterprise plan allows CrewAI cloud, a dedicated VPC, NAT, or the customer's own infrastructure, and mentions SAM certified and Fed Ramp High options without a supporting page. The privacy policy refers to service providers for hosting, analytics and email delivery but never names them, and no subprocessor list is published. CrewAI, Inc. is a Delaware corporation and the terms are governed by New York law, which places the legal centre of gravity in the United States. The domain resolves to a United States address on an Amazon anycast node, but that reflects site delivery, not where customer data lives. A Vanta trust centre exists at trust.crewai.com; it renders nothing without JavaScript, so none of this could be confirmed from it.
Things to keep in mind
Risks and trade-offs to weigh before adopting CrewAI.
- The terms let CrewAI use your data, including everything your agents produce, to train and develop its models, and nothing on the site documents a way to opt out
- Any feedback you send is licensed perpetually, irrevocably and worldwide, and is explicitly not treated as confidential
- The vendor itself warns that outputs may repeat what other customers received, may carry biases from training data, and may be discriminatory or offensive
- Autonomous agents acting inside your CRM, mailbox, files and payment systems widen the blast radius of a single bad instruction far beyond a chatbot's
- Connecting Gmail, Drive, Contacts or Calendar through OAuth exposes personal data belonging to third parties who never agreed to it
- Human approval gates are presented as a feature to configure rather than a default, which makes it easy to automate away the review step precisely where judgment matters
- With no hosting location, no subprocessor list and no retention rule for platform data published, you cannot assess where your data goes or how long it stays
Setup & Integrations
Technical difficulty
Two very different paths. The no-code route is easy: create an account, open Crew Studio, and build from prompts, a canvas and ready-made workflow templates with an AI copilot alongside. The code route is a developer task: install through uv, configure model API keys, set up the CLI, and define agents in YAML or Python. Two things trip people up there — the hierarchical process will not work unless a manager LLM is configured, and getting good behaviour means tuning attributes such as max_rpm, max_iter and allow_delegation. Enterprise customers get a 45-day onboarding programme.
Deployment
Integrations
Supported languages
Behind CrewAI
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Is CrewAI free to use?
Do I need to know how to code?
What is the difference between the sequential and hierarchical processes?
What kinds of memory do agents have?
Where can I deploy my agents?
Is there an API?
Which tools does CrewAI connect to?
Is my data used to train models?
How long is my data kept?
Is there a minimum age?
Should you pick CrewAI?
CrewAI's real strength is continuity. The same product covers an open source Python framework, a no-code editor whose workflows export to Python, and a control plane that watches every run in production. That removes the break most teams hit between a business prototype and something engineers are willing to operate, and it explains why the customer list runs to PwC, IBM, AWS, Docusign and Gelato rather than side projects. Observability and governance are not add-on modules here: the Control Plane sits in the execution path itself, and tracing, cost accounting, access control, audit trails and human approval gates come with it.
The reservations are just as clear, and they are mostly about what the site does not say. Nothing above the free tier carries a price. The GDPR is never mentioned anywhere, no data processing agreement or subprocessor list is published, and no hosting country or region is disclosed — the trust centre that might settle all of this renders nothing without JavaScript. Meanwhile the terms of use grant CrewAI an explicit licence to use customer data, agent outputs included, to train and develop models, and no opt-out is documented. For a platform pitched at regulated, mission-critical environments, that gap between the governance promised inside the product and the transparency offered outside it is worth raising before signing.
The free plan is an evaluation tool, not a production one: fifty executions a month with no overage settles that. CrewAI makes most sense for organisations that already have agent pilots running and a scaling and governance problem to solve, and for engineering teams who want a framework rather than a closed product. Anyone who needs a mobile app, a published price, or documented European data commitments should look elsewhere for now.
- Choosing a selection results in a full page refresh.
- Opens in a new window.