CrewAI

Model

1. TouchesModel APIs: Touches

   Model APIs, touches. Agents call models as provider/model-id through native SDKs for OpenAI, Anthropic, Gemini, Azure, Bedrock, and Snowflake Cortex, with the team's own API keys; model access stays with those platforms.

2. Model gateway: Touches

   Model gateway, touches. Every provider without a native integration runs through LiteLLM, and each agent can name its own model; routing, fallbacks, and spend limits across providers stay with a gateway.

Build

1. CrewAICoreOrchestration: Core

   Orchestration, core. This is CrewAI's primary role: Crews of agents with roles, goals, and tools run tasks, and Flows add start, listen, and router steps, @persist state, checkpoints, and @human_feedback review, in Python 3.10 to 3.13.

2. SupportsTool use: Supports

   Tool use, supports. Agents use built-in and custom tools and MCP servers, and AMP adds connectors for Slack, Gmail, Salesforce, Jira, and other apps with triggers; the systems those tools reach stay with the team.

Data

1. Retrieval: Supports

   Retrieval, supports. memory=True gives a crew shared memory stored in LanceDB and embedded with OpenAI text-embedding-3-large by default; indexing company data at scale stays with a retrieval product.

Evaluate

1. Evaluation: Touches

   Evaluation, touches. AMP lists LLM testing, training, and a hallucination guardrail; datasets, graders, and experiments stay with an evaluation tool such as Braintrust or Arize, which integrate with CrewAI.

Operate

1. Observability: Supports

   Observability, supports. AMP traces every crew and flow execution and exports traces and logs to your own OpenTelemetry collector or Datadog; the library also works with Langfuse, Arize Phoenix, and Braintrust.

2. Cloud deployment: Supports

   Cloud deployment, supports. AMP deploys crews and flows from GitHub to managed infrastructure with automatic scaling, and Enterprise runs on CrewAI cloud, a dedicated VPC, or your own infrastructure; Basic caps runs at 50 executions a month.

3. Guardrails: Supports

   Guardrails, supports. CrewAI task guardrails validate and transform task outputs before they reach the next task. Failed validation triggers the configured retry limit.