CrewAI Review: The Open-Source Framework and the AMP Platform

CrewAI is two things under one name. The first is an open-source Python framework, MIT-licensed, where you define AI agents with a role, a goal and tools, group them into crews and wire crews into event-driven Flows that run on your own machines. The second is CrewAI AMP (Agent Management Platform), the paid layer that builds, deploys, triggers and monitors those crews. AMP's free Basic plan includes 2 automations and 50 workflow executions a month, with no additional executions, while Enterprise is priced on request. Our verdict is simple. CrewAI is a strong pick for Python teams that want multi-agent prototypes fast, and a weaker one when nobody on staff writes code.

Two products share the CrewAI name

The framework is the Python package crewai, so you can install it, change it and ship it inside your own product without paying CrewAI. It needs Python 3.10 or newer and below 3.14, installs its CLI with uv tool install crewai, and runs crews locally with crewai run. Projects made with crewai create crew are JSON-first: agents live in agents/*.jsonc files and crew settings in crew.jsonc, while the older Python and YAML layout still works behind the --classic flag. The package does not depend on LangChain, although agents can still use LangChain and LlamaIndex tools.

AMP is the hosted side, made by CrewAI, Inc. Every crew or Flow you deploy there becomes an automation with its own API endpoint, triggers, traces and access rules. You deploy from GitHub, any Git provider, a ZIP upload, the CLI or Studio, the no-code builder. AMP runs in CrewAI's cloud, as a self-hosted install called CrewAI Factory, or as a hybrid of the two. Keep the two products apart when you budget. The framework never sends an invoice. AMP is where plans, execution caps and governance controls live, and our guide to whether CrewAI is free lists the features that need an AMP account.

How crews, Flows and agents fit together

An agent in CrewAI has a role, a goal, a backstory, tools and an optional model of its own. Tasks give it work: each needs a description and an expected output, and can return structured data. A crew runs those tasks in one of two processes. Sequential is the default, a fixed order in which each output feeds the next task. Hierarchical adds a manager, set through manager_llm or manager_agent, that plans, delegates, reviews and validates. The docs advise starting sequential. Delegation is off by default for good reason, since agents allowed to delegate can pass work back and forth indefinitely.

Flows are the control layer. They chain crews, single agents and plain Python into event-driven steps with shared state, using the @start, @listen and @router decorators plus conditions, loops and branching. @persist saves state to local SQLite, and @human_feedback pauses a flow until a person approves, rejects or asks for a revision. CrewAI's own guidance is to use a crew when the path to the answer is open-ended and a Flow when you control the steps and their order. Its docs recommend starting production apps with a Flow that runs crews inside its steps.

Around that core sit the pieces buyers ask about most. The tool catalog lists 75+ tools for search, scraping, files, databases and cloud storage, and MCP servers plug straight into an agent. Memory is one unified class, rebuilt in March 2026 on LanceDB. Checkpointing, added in April 2026, lets a failed run resume or fork. Native model connections cover OpenAI, Anthropic, Gemini, Azure, Bedrock and Snowflake Cortex. Other providers run through the optional LiteLLM extra, and local Ollama models work through that extra or as an OpenAI-compatible endpoint.

What AMP's two plans include

AMP has exactly two plans, and the pricing page prints a dollar figure for neither: Basic is listed as Free and Enterprise as Custom, behind a Request trial button ^[1]. Because no paid price appears, no billing period appears either. Basic is not a trial. It includes 2 automations and 50 workflow executions a month on CrewAI's cloud, and 50 is also the ceiling, since additional executions are not included on Basic. By our arithmetic, that is about 1.7 executions a day across a 30-day month, so size any pilot to that ceiling.

The split is generous on building and thin on governance. Studio, private agent and tool repositories, workflow templates, GitHub integration, cron scheduling, human-in-the-loop input, workflow chat in Slack or Teams and export as an MCP server are all on Basic. Enterprise adds enterprise connectors, a dedicated VPC and NAT, SSO, role-based access control, PII redaction for traces and its policies, and a 45-day onboarding. The support rows list training on Enterprise; the plan card adds forward-deployed engineering and training a la carte. CrewAI's Terms say fees are set in an Order Form, payments are non-cancelable and fees paid are non-refundable ^[2]; our CrewAI pricing guide covers renewal notice.

Costs and defaults that surprise newcomers

Open source is not zero cost. Every agent step calls your model provider on your own API key, and AMP needs your keys too, so token bills come from OpenAI, Anthropic or whichever provider you pick. The default model is an OpenAI one, and default memory and knowledge embeddings also call OpenAI unless you set another embedder. Many tools need third-party API keys billed by those services, Tavily, Exa and Firecrawl among them, while Serper has a free tier. Hosting a self-run crew is yours too. A fully local path does exist, with Ollama serving both the model and the embedder.

Token growth is the second surprise. Multi-agent runs make many model calls, and an agent can loop on a tool. Since October 2026, the fifth identical tool call with the same arguments and result stops the task, and that limit cannot be changed. Spend controls cover the rest: max_iter, max_execution_time and max_rpm on agents, and max_tokens on tasks. On AMP, Cost Limit policies email recipients when a daily, weekly or monthly budget is crossed. They never stop a run.

Telemetry is on by default. The open-source package sends anonymous usage data that includes tool names and agent roles, so keep personal information out of both . Setting CREWAI_DISABLE_TELEMETRY or OTEL_SDK_DISABLED to true turns it off. The share_crew option is a separate opt-in that sends goals, backstories, task descriptions and outputs to CrewAI to improve the library and train models. Tracing is a third switch, independent of both. On AMP, the Terms let CrewAI use Customer Data, outputs included, to improve its services and train models, and they print no opt-out. Its Privacy Policy separately says CrewAI never uses Google user data for advertising or for training AI models.

Two changes and one limit round this out. CodeInterpreterTool was removed in April 2026 and allow_code_execution is deprecated, so code now runs in E2B or Daytona sandbox tools, each with its own account . Tool-result caching became opt-in in July 2026, whatever older pages say. Studio export runs one way. Downloaded code cannot be imported back, and redeploying edited code creates a separate automation without the visual editor. On certifications, CrewAI says (May 2025) that CrewAI Factory has a SOC 2 Type II report and a HIPAA Security Rule gap assessment.

How we score CrewAI

We score every product on the same eight axes, each read from its own evidence, and weight capability and AI depth above support once a product runs reliably. User sentiment is one axis of the eight, never the whole verdict. For CrewAI, the axes that move the result most are capability, where the framework and Flows are deep, and value, where a free framework meets model bills that only you can size. The full breakdown is below.

That composite lands CrewAI at 4.1/5. It credits a framework that costs nothing to run yourself, a broad tool catalog, Flows with real state and human review, and a free platform tier that already includes Studio and tracing. It holds back for the token bills that grow with every agent, governance controls that sit only on Enterprise, and a pricing page that leaves paid costs to a sales conversation. Developers gain the most here. Business teams without Python skills lean on Studio, which still needs an LLM connection on their own provider key.

Reviewers on G2 mostly praise how CrewAI splits a complex job into role-based agents and how fast multi-agent prototypes come together. The recurring complaints are token use and run time that grow quickly with multi-agent workflows, and agents that can loop on tools. One reviewer adds that production use needs more work on guardrails, state and error handling. Opinions on pricing split, with some finding automation pricing high. That pattern matches the product: quick to start, heavier to run at scale, and only as cheap as the models behind it.

Is CrewAI the right first step for your agent project?

Start with the open-source framework if your team writes Python and wants role-based agents running this week: install it, build a sequential crew, wrap it in a Flow and watch the token meter before anything else. Add AMP's Basic plan when you need deployment, triggers and traces, and treat its 50 workflow executions a month as a pilot budget. Move to Enterprise talks once you need more than 50 workflow executions a month or more than 2 automations, enterprise connectors, or SSO, role-based access or your own VPC. If explicit graph control matters more than ready-made roles, read our CrewAI vs LangGraph comparison first.