CrewAI

The framework ships a large tool library and custom tools and connects to MCP servers as tools, and CrewAI AMP adds authenticated enterprise integrations such as Salesforce, HubSpot, Slack, Gmail, Microsoft 365, Jira, Notion, Stripe and Zendesk.

Sourcedocs-platform.crewai.com/platform/en/features/tools-and-integrationsread 2026-09-27

Flows are event driven workflows with managed state and deterministic control, and Crews are teams of role based agents that delegate to one another; a Flow step can hand off to a Crew, processes run sequential or hierarchical with a manager agent, tasks can be conditional, and checkpointing lets runs be replayed or forked.

Sourcedocs.crewai.com/en/concepts/flowsread 2026-09-27

Knowledge sources give agents a reference library of the customer's files and data, embedded into a vector store (ChromaDB by default, Qdrant supported) through a provider neutral RAG client and queried during tasks; new sources are added without retraining.

Sourcedocs.crewai.com/en/concepts/knowledgeread 2026-09-27

Flows pause at @human_feedback steps for a person's review, and CrewAI AMP routes those requests by email to named reviewers with fallback responses on timeout; tasks can require human input before they complete, and a hallucination guardrail checks outputs.

Sourcedocs-platform.crewai.com/platform/en/features/flow-hitl-managementread 2026-09-27

CrewAI offers RBAC with feature and entity level permissions and custom roles, SSO through WorkOS, Microsoft Entra ID, Okta, Auth0 or Keycloak on both SaaS and self hosted Factory, a secrets manager with cloud workload identity federation, and PII redaction for traces. No certification is named in the docs, and the trust center sits at trust.crewai.com.

Sourcedocs-platform.crewai.com/platform/en/features/rbacread 2026-09-27

Traces in CrewAI AMP record every execution from inputs to outputs, including agent reasoning, task details, tool calls and outputs, tokens, timings and cost estimates, with OpenTelemetry export and a Datadog integration; the framework also emits events and integrates with tracing tools.

Sourcedocs-platform.crewai.com/platform/en/features/tracesread 2026-09-27

A unified Memory class stores what agents learn in a hierarchical scope tree that can be inspected, recalls by semantic similarity, recency and importance with a configurable recency half life, and forgets by scope, so one project's or customer's memory can be deleted without wiping the rest; it works with crews, agents and flows.

Sourcedocs.crewai.com/en/concepts/memoryread 2026-09-27

Enterprise deploys on CrewAI cloud, a dedicated VPC or the customer's own infrastructure (CrewAI Factory, self hosted), and the MIT framework runs anywhere the customer runs Python.

Sourcecrewai.com/pricingread 2026-09-27

Every plan lists agentic workflow templates, and the AMP Marketplace offers reusable assets and templates to install, but the templates and what each one does are not listed. The open source examples and cookbooks are starting points a developer adapts.

Sourcedocs-platform.crewai.com/platform/en/features/marketplaceread 2026-09-27

Events in Gmail, Google Calendar and Drive, Outlook, OneDrive, Microsoft Teams, Slack, HubSpot, Salesforce and Zapier can launch the chosen crew or flow through CrewAI AMP triggers, and webhook automation starts runs from outside systems.

Sourcedocs-platform.crewai.com/platform/en/guides/automation-triggersread 2026-09-27

Each agent can take its own model, chosen by the developer from many providers through their native SDKs or any LLM connection, including custom LLM implementations, and AMP supports Azure OpenAI and Vertex AI setups.

Sourcedocs.crewai.com/en/concepts/llmsread 2026-09-27

Deployed crews expose a REST API (GET /inputs, POST /kickoff, POST /resume, GET /status), the AMP platform API manages automations, users, roles and service accounts with access tokens, the Python framework and CLI are open source, and crews can be exported as MCP servers.

Sourcedocs-platform.crewai.com/api/introductionread 2026-09-27

crewai test runs a crew for a set number of iterations and scores each task from 1 to 10 with an evaluator model, and AMP lists LLM testing and agent training that turns runs into improvement data; evaluation integrations such as Patronus are documented.

Sourcedocs.crewai.com/en/concepts/testingread 2026-09-27

Agents reach the browser through third party engines added as tools, among them Stagehand, Browserbase, Hyperbrowser, MultiOn and a Selenium scraper, plus scraping tools, and CrewAI runs no browser of its own.

Sourcedocs.crewai.com/en/tools/web-scraping/overviewread 2026-09-27