CrewAI is an open source Python framework from crewAI, Inc. for orchestrating autonomous AI agents, built on two ideas: Crews, which are teams of role-playing agents, and Flows, which are event-driven workflows that hold state and control the order of work. The name also covers CrewAI AMP, the company's commercial Agent Management Platform for deploying and governing those crews and flows. This page explains what each part does, which version is current, how models are connected, what the plans include, what the framework sends as telemetry and how the release line has moved since 1.0. It is written for developers, architects and the people who approve frameworks for a team. Every fact comes from CrewAI's own documentation, repository, changelog and pricing page. Last checked: 2026-11-15.
This is an independent guide from NHI Mgmt Group, based on CrewAI's public documentation as of November 2026. It is not sponsored or reviewed by CrewAI. Product features change, so confirm settings against CrewAI's current documentation.
Key takeaways
- Two building blocks. Flows are the "manager" or process definition of an application, and Crews are the teams of agents a Flow hands work to. CrewAI's advice for production is to start with a Flow and use a Crew inside a Flow step.
- Fast release cadence. Version 1.0.0 shipped on 18 October 2025 and the changelog reached 1.15.41 on 13 November 2026, with patch releases often several times a week.
- The framework is free and MIT licensed. The repository LICENSE carries the MIT text with "Copyright (c) 2025 crewAI, Inc.", and the package needs Python 3.10 or later and below 3.14.
- AMP is the paid layer. The pricing page lists a Free plan with 50 executions a month and two automations, and a custom-priced Enterprise plan that adds VPC or self-managed deployment, SSO and RBAC.
- Telemetry is on by default and can be switched off. Anonymous usage data is collected unless
CREWAI_DISABLE_TELEMETRYis set, and AMP tracing has its own switch,CREWAI_TRACING_ENABLED=false.
What is CrewAI?
CrewAI's documentation calls it "the leading open-source framework for orchestrating autonomous AI agents and building complex workflows". That is the company's own description. In practice the framework gives you four objects to work with:
- Agents: role-playing specialists, each with a role, a goal and a set of tools. They decide what to do based on their role and the tools available to them.
- Tasks: units of work assigned to agents according to their capabilities, each with a description and an expected output.
- Crews: groups of agents and tasks that run under a process. The
Processenum has two values:Process.sequential, where tasks run in list order and each output can feed the next, andProcess.hierarchical, where a manager plans, delegates, reviews outputs and judges completion. A hierarchical crew needs eithermanager_llmor a custommanager_agent. - Flows: the application backbone. Flows provide state management, event-driven execution and control flow such as conditionals, loops and branching, and they can call crews, single agents or plain Python.
The project is led by crewAI, Inc. The pyproject.toml names Joao Moura as author. The README says more than 100,000 developers have been certified through the company's community courses at learn.crewai.com.
| Building block | What it controls | Typical use |
|---|---|---|
| Flow | State, order of steps, branching, persistence, human feedback | The production application and its control logic |
| Crew (sequential) | A fixed task order across several agents | Repeatable pipelines such as research then write |
| Crew (hierarchical) | A manager agent that delegates and reviews | Open-ended work where the split of tasks is not known in advance |
| Single agent | One role, goal and tool set | A focused step inside a Flow |
| AMP automation | A deployed Crew or Flow with its own API, triggers, traces and access rules | Running and governing agents in production |
Key facts at a glance
As of 15 November 2026, from CrewAI's documentation, repository, changelog and pricing page.
| Field | CrewAI |
|---|---|
| Owner | crewAI, Inc. |
| Type | Open source multi-agent framework, plus a commercial platform (CrewAI AMP) |
| Language | Python |
| Python requirement | 3.10 or later and below 3.14 (requires-python = ">=3.10, <3.14") |
| Licence | MIT (repository LICENSE, Copyright (c) 2025 crewAI, Inc.) |
| First stable release | 1.0.0, 18 October 2025 |
| Latest release in the changelog | 1.15.41, 13 November 2026 |
| Package manager | uv (uv tool install crewai installs the CLI) |
| Packages in the repository | crewai, crewai-core, crewai-cli, crewai-tools, crewai-files |
| Native model integrations | Anthropic, AWS Bedrock, Azure, Google (Gemini API), OpenAI, Snowflake Cortex; others through the LiteLLM extra |
| MCP support | Yes: mcps field on agents and MCPServerAdapter in crewai-tools |
| Price | Framework free; AMP Free plan and custom Enterprise plan |
| Documentation | docs.crewai.com (framework) and docs-platform.crewai.com (AMP) |
Plans and pricing
The open source framework costs nothing. CrewAI AMP is priced per plan. The table below copies crewai.com/pricing as of 15 November 2026. The page does not publish dollar figures for Enterprise, a seat count for either plan, or an overage rate.
| Item | Free | Enterprise |
|---|---|---|
| Price | Free | Custom |
| Executions included per month | 50 | Sized to the workflow |
| Overage | Not stated by CrewAI | "Flexible overage" (rate not stated by CrewAI) |
| Automations | 2 | Custom |
| Deployment | CrewAI cloud | CrewAI cloud, your own VPC or your own infrastructure |
| Dedicated VPC and NAT | Enterprise plan feature | Included |
| Sign-in | Not stated by CrewAI | SSO (Microsoft Entra, Okta) |
| Access control | Enterprise plan feature | Role-based access control |
| Connectors | Standard tools and triggers | Standard tools and triggers plus enterprise connectors |
| Support | Community | Dedicated support, Slack or Teams channel, on-site help |
| Onboarding | Not stated by CrewAI | Training, onboarding (45 days), deployment and development help |
Features on both plans: the pricing page lists the Studio visual editor, private agent and tool repositories, workflow templates, GitHub integration, MCP server and UI component export, tracing, OpenTelemetry export, guardrails, human-in-the-loop input, workflow chat in the UI and in Slack or Teams, a usage dashboard with token counts and performance metrics, automatic scaling and cron scheduling.
Availability: SSO, RBAC, dedicated VPC and NAT, self-managed deployment and enterprise connectors are listed on the Enterprise plan only. Exact Enterprise pricing is available on request through the "Request Info" route on the pricing page.
Models available
CrewAI does not ship its own model. You choose one per agent, per crew manager or per LLM object, always with a provider prefix in the form provider/model-id. The LLM documentation, as of 15 November 2026, describes two routes:
| Route | Providers named in the docs (alphabetical) | Install |
|---|---|---|
| Native SDK integration | Anthropic, AWS Bedrock, Azure, Google (Gemini API, and Vertex AI through the same Gemini integration), OpenAI, Snowflake Cortex; DeepInfra through the native OpenAI client | Included with crewai (provider SDK extras where needed) |
| LiteLLM | Cerebras, Fireworks AI, Groq, Hugging Face, IBM watsonx.ai, Meta-Llama, Mistral, Nebius AI Studio, NVIDIA NIM, Ollama, OpenRouter, Perplexity AI, SageMaker, SambaNova | uv add 'crewai[litellm]' |
There are three ways to set a model. You can set MODEL=provider/model-id in the environment or the project .env file, alongside the provider key such as OPENAI_API_KEY or ANTHROPIC_API_KEY. You can set an agent's llm: field in its configuration file. Or you can create LLM(model="provider/model-id") in code with options such as timeout, max_tokens and response_format. Release 1.15.27 (9 October 2026) added DeepInfra as an OpenAI-compatible provider, and 1.15.28 (14 October 2026) began taking context window sizes from LiteLLM data.
Features and modes
Flows
Flows are plain Python classes driven by decorators:
@start(): marks an entry point. A flow can have several, and a start can be gated on another method, a router label or a callable condition.@listen(): runs a method when another method finishes.@router(): sends execution down a branch based on the return value. Branch labels are declared with aLiteral[...]return type oremit=.or_()andand_(): trigger a listener when any, or all, of the listed methods have produced output.- State: unstructured (a dictionary-style
self.state) or structured (a Pydantic model passed asFlow[ExampleState]). Every state gets anidautomatically. @persist: saves state across restarts, with SQLite as the default backend.kickoff(inputs={"id": ...})resumes a run, andkickoff(restore_from_state_id=...)forks a new run from a snapshot.@human_feedback: pauses a flow for human input (CrewAI 1.8.0 or later). Withemit=, a model maps free-form feedback to one of the listed outcomes.
Crews, tools and MCP
Crews combine agents, tasks and a process as described above. Tools come from the crewai-tools package or your own code. MCP servers connect in two ways:
mcpsfield on an agent (recommended in the docs): accepts a remote URL (optionally with#tool_nameto pick one tool), a connected integration slug, or structured configs fromcrewai.mcp:MCPServerStdio,MCPServerHTTP(streamable by default) andMCPServerSSE. Tool filters includecreate_static_tool_filter(allow and block lists) andcreate_dynamic_tool_filter.MCPServerAdapter: for manual connection management, used as a context manager, withconnect_timeoutdefaulting to 30 seconds. Install it withuv pip install 'crewai-tools[mcp]'. The docs note that the adapter mainly supports MCP tools, not prompts or resources.
Recent additions
Changelog entries from autumn 2026 show where the framework is going: experimental conversational Agent sessions (1.15.36, 5 November 2026), spreadsheet reading for agents in crewai-files and scaffolding for standalone agent projects (1.15.37, 6 November), a crewai skill search command (1.15.34, 1 November) and context compaction for conversational flows (1.15.31, 27 October).
Supported IDEs and platforms
CrewAI is a Python library, so it runs wherever a supported Python version runs. The tooling is the crewai CLI, installed with uv:
| Command | What it does |
|---|---|
uv tool install crewai | Installs the CLI globally (--upgrade to update) |
crewai create crew <name> | Generates a project |
crewai install | Installs project dependencies |
crewai run | Runs the crew from the project root |
The default scaffold is now JSON-first: crew.jsonc, an agents/ folder with .jsonc agent files, plus knowledge/, skills/ and tools/ folders, a .env file and pyproject.toml. The older layout with crew.py and YAML files under config/ is still available through --classic.
For hosting, CrewAI AMP deploys an automation by publishing directly from Studio, or from code through GitHub, another Git provider, a ZIP upload or the CLI. The README describes on-premise and cloud deployment options for AMP, and the pricing page places your own VPC and your own infrastructure on the Enterprise plan.
Enterprise and admin controls
Admin controls sit in CrewAI AMP rather than in the library. The AMP documentation groups its features into three areas:
| Area | Features (AMP documentation) |
|---|---|
| Build | Studio, Deploy from Code, Tools and Integrations, Agents Repository |
| Operate | Automations, Triggers and API, Traces, Webhook Streaming |
| Govern | Agent Control Plane, Role-Based Access Control, Secrets Manager, PII Trace Redactions |
Each deployed Crew or Flow becomes an automation with its own API, triggers, traces and access rules. The Enterprise plan adds SSO through Microsoft Entra or Okta, RBAC and dedicated VPC networking. Identity design for agents, tool credentials and approval hooks is covered in the CrewAI security guide, so it is not repeated here.
Data handling and privacy
CrewAI's telemetry page sets out what the framework sends by default and how to stop it:
| Data | Default | Control |
|---|---|---|
| CrewAI and Python versions, random crew IDs, process type, agent and task counts, tool names, model name and temperature, task timings and success or failure (exception class name only) | Collected | CREWAI_DISABLE_TELEMETRY=true |
Execution environment signals (coding assistant if any, CI or container, project_id, coarse CPU band) | Collected | CREWAI_DISABLE_TELEMETRY=true |
| Agent goals and backstories, task descriptions and context, crew inputs and outputs, final output | Not collected | Sent only if share_crew=True |
| All OpenTelemetry instrumentation and CrewAI's exporter | On | OTEL_SDK_DISABLED=true (does not stop AMP tracing) |
| AMP tracing | Per configuration | CREWAI_TRACING_ENABLED=false or tracing=False on the Crew or Flow |
The documentation states that by default "NO personal data is collected" concerning prompts, task descriptions, responses or secrets. Accepted values for the switches are true, 1, yes or on in any case, and unrecognised values leave telemetry on. The docs warn that share_crew data may include personal information if you put it in agent or task content, so review it for privacy compliance first. What a model provider retains is governed by that provider's own terms, not by CrewAI.
Release history
Dated entries from CrewAI's changelog, as of 15 November 2026.
| Date | Release | What changed |
|---|---|---|
| 18 Oct 2025 | 1.0.0 | First stable release |
| 20 Oct 2025 | 1.1.0 | Minor release |
| 21 Oct 2025 | 1.2.0 | Minor release |
| 27 Oct 2025 | 1.3.0 | Minor release |
| 1 Nov 2025 | 1.4.0 | Minor release |
| 7 Nov 2025 | 1.5.0 | Minor release |
| 9 Oct 2026 | 1.15.27 | DeepInfra as OpenAI-compatible provider; per-run cost and time tracking in eval |
| 14 Oct 2026 | 1.15.28 | crewai eval failure reporting; LiteLLM-based context windows |
| 23 Oct 2026 | 1.15.29 | CrewAgentExecutor removed (breaking); Azure DocumentDB vector search |
| 27 Oct 2026 | 1.15.31 | Conversational flow context compaction; path traversal fix in output_file |
| 1 Nov 2026 | 1.15.34 | crewai skill search; planning configuration docs |
| 5 Nov 2026 | 1.15.36 | Experimental conversational Agent sessions; prompt for missing API keys on crewai run |
| 11 Nov 2026 | 1.15.40 | chromadb removed from the embedding function layer; standalone agents scaffolded as JSON |
| 13 Nov 2026 | 1.15.41 | Approved private URL hosts; earlier step outputs kept across human-feedback resumes |
The changelog does not publish a support or long-term support policy for older minor versions as of November 2026, so the practical rule is to track the latest release and read each entry before upgrading.
Limits and known constraints
- Python ceiling: the package supports Python below 3.14, so check the interpreter in base images before upgrading the runtime.
- Breaking changes inside 1.x: 1.15.29 removed
CrewAgentExecutor. Pin the version in your lock file and read the changelog before taking a new patch. - Scaffold change: new projects default to the JSON layout. Teams with existing
crew.pyprojects can keep them, and use--classicfor new ones that should match. - MCP adapter scope:
MCPServerAdapterfocuses on MCP tools and the primary text output, so prompts, resources and multi-modal results need custom handling. - Experimental features: conversational Agent sessions are labelled experimental as of November 2026.
- Hierarchical process setup: a hierarchical crew will not run without
manager_llmormanager_agent.
Security guidance
This page covers what CrewAI is, how it is versioned and what the platform includes. Hardening is covered in the CrewAI security guide, which goes through agent identity, tool permissions and hooks, MCP servers, secrets and logging. From the framework docs, the customer-side settings to decide early are: whether telemetry and tracing stay on (CREWAI_DISABLE_TELEMETRY, CREWAI_TRACING_ENABLED), keeping share_crew off unless reviewed, keeping provider keys in .env or a secrets manager rather than in agent files, filtering MCP tools with static allow lists, and following the MCP page's advice to trust a server before use and, for SSE, to validate Origin headers and bind local servers to 127.0.0.1. Apply CrewAI updates promptly. The wider discipline is covered in the Securing Agent Frameworks Guide.
Practitioner checklist
- Confirm Python 3.10 to 3.13 in every build image and pin
crewaiin a lock file. - Read the changelog before each upgrade, especially for removals such as
CrewAgentExecutorin 1.15.29. - Structure production apps as a Flow with crews inside steps, and use structured Pydantic state where data shape matters.
- Set
CREWAI_DISABLE_TELEMETRYandCREWAI_TRACING_ENABLEDexplicitly in each environment so data sharing is a deliberate choice. - Keep
share_crewoff unless a privacy review has approved the data it sends. - Set models with explicit
provider/model-idvalues and keep provider keys out of source control. - Use static tool filters on MCP connections and only connect MCP servers you trust.
- Check which AMP features you need (SSO, RBAC, VPC, self-managed hosting) against the plan that lists them before committing.
Standards and references
- CrewAI: Introduction
- CrewAI: Changelog
- CrewAI: Installation
- CrewAI: Flows
- CrewAI: LLMs
- CrewAI: Telemetry
- CrewAI: AMP introduction
- CrewAI: Pricing
Related NHI Mgmt Group resources: CrewAI security guide · Securing Agent Frameworks Guide · Agent Frameworks vs Workflow Orchestration · Agent Frameworks vs Managed Agent Platforms