#1Top pickFree option
Open-source framework for building stateful, controllable agents as graphs
LangGraph is an open-source framework from LangChain for building agents as graphs of steps with state, memory and human-in-the-loop checkpoints. It gives developers fine control over how an agent moves between tools and decisions. It is MIT licensed, with LangSmith available for tracing and deployment.
Why it's here
- +Fine control over agent flow
- +Built-in persistence and human approval steps
- +Large ecosystem and docs
Watch out for
- −More code than higher-level frameworks
- −Concepts take time to learn
- −Best tooling ties into LangSmith
#2Free option
Open-source Python framework for teams of role-based AI agents
CrewAI is an open-source Python framework for building multi-agent systems where agents with defined roles work together on tasks. It is known for being quick to start with. CrewAI also runs a hosted platform with a free Basic plan and custom Enterprise pricing.
Why it's here
- +Simple, readable agent definitions
- +MIT licensed
- +Large community and examples
Watch out for
- −Less fine-grained control than LangGraph
- −Multi-agent runs can be costly in tokens
- −Debugging agent hand-offs is hard
O#3Free option
OpenAI's lightweight open-source SDK for building agents with tools and handoffs
The OpenAI Agents SDK is an open-source framework for building agents that use tools, hand off to other agents and apply guardrails, with tracing built in. It is small and Python-first, with a TypeScript version too. It is MIT licensed and works with OpenAI and other model providers.
Why it's here
- +Few concepts to learn
- +Built-in tracing and guardrails
- +MIT licensed
Watch out for
- −Fewer built-in integrations than LangChain
- −Best experience with OpenAI models
- −Less suited to complex graph workflows
#4Free option
Anthropic's SDK for building agents on the same harness that powers Claude Code
The Claude Agent SDK is an open-source SDK from Anthropic for building agents that use the same tools and agent loop as Claude Code, including file access, command execution, subagents and MCP. It is available for Python and TypeScript. The SDK is MIT licensed; you pay for Claude API usage.
Why it's here
- +Battle-tested agent loop from Claude Code
- +Built-in tools and MCP support
- +Python and TypeScript
Watch out for
- −Designed around Claude models
- −API costs for long sessions
- −Less suited to visual, no-code builders
#5Free option
Type-safe Python agent framework from the team behind Pydantic
Pydantic AI is an open-source Python framework for building agents with typed inputs, outputs and tools, made by the Pydantic team. It works with many model providers and pairs with Pydantic Logfire for observability. It is MIT licensed.
Why it's here
- +Typed, validated outputs
- +Model-agnostic
- +Feels like normal Python
Watch out for
- −Python only
- −Fewer prebuilt integrations
- −Multi-agent patterns are more manual
#6Free option
TypeScript framework for AI agents, workflows, memory and evals
Mastra is a TypeScript framework for building AI agents, workflows and AI-powered apps, with memory, tools and observability included. It suits JavaScript and TypeScript teams that want to stay in their own stack. The core is Apache-2.0 licensed, with some enterprise features under a separate license.
Why it's here
- +Native TypeScript
- +Workflows, memory and evals in one package
- +Local dev playground
Watch out for
- −Not for Python teams
- −Younger ecosystem
- −Enterprise features under a different license
#7Free option
Microsoft's open-source agent SDK for .NET and Python, successor to AutoGen
Microsoft Agent Framework is an open-source SDK for building AI agents and multi-agent workflows in Python and .NET. It brings together ideas from AutoGen and Semantic Kernel, and Microsoft recommends it for new projects instead of AutoGen, which is now in maintenance mode. It is MIT licensed.
Why it's here
- +First-class .NET and Python support
- +Backed by Microsoft
- +Clear migration path from AutoGen
Watch out for
- −Younger than LangGraph or CrewAI
- −Best fit leans toward Azure
- −Docs still catching up
#8Free option
Google's open-source toolkit for building and deploying multi-agent systems
Google Agent Development Kit (ADK) is an open-source framework for building, testing and deploying AI agents and multi-agent systems. It is optimized for Gemini and Google Cloud but works with other models. It is Apache-2.0 licensed.
Why it's here
- +Built-in eval and dev UI
- +Easy deploy to Google Cloud
- +Apache-2.0 licensed
Watch out for
- −Best fit with Gemini and Google Cloud
- −Smaller community than LangGraph
- −Fast-moving APIs
#9Free option
Open-source framework for agents over your documents and data
LlamaIndex is an open-source framework for building agents and retrieval apps that work over your documents and data, with Workflows for multi-step agents. The company also sells LlamaParse, a document parsing and extraction service. The framework is MIT licensed; LlamaParse has a free credit tier and paid plans.
Why it's here
- +Strong document and retrieval tooling
- +Many data connectors
- +MIT licensed
Watch out for
- −Best parsing is a paid service
- −API surface changes often
- −Less focused on general agents
#10Free option
Open-source library and cloud for AI agents that control web browsers
Browser Use is an open-source Python library that lets AI agents operate a web browser, plus a cloud service that runs those browsers for you. Developers use it to automate web tasks on sites without APIs. The library is MIT licensed; the cloud is pay-as-you-go with a $5 minimum top-up.
Why it's here
- +Large open-source community
- +Works with many models
- +Hosted browsers when you need scale
Watch out for
- −Requires coding
- −Websites change and break flows
- −Model and proxy costs add up
#11Free option
Open-source platform for building AI apps, agents and RAG workflows
Dify is an open-source platform for building AI applications, agents and retrieval workflows with a visual editor. You can self-host the Community Edition with Docker or use Dify Cloud. The cloud Sandbox plan is free with 200 message credits and Professional costs $590 per workspace per year.
Why it's here
- +Self-hostable
- +Visual workflow and RAG tools
- +Works with many model providers
Watch out for
- −License has extra conditions beyond Apache-2.0
- −Self-hosting needs maintenance
- −Less suited to non-technical users
Free · from $49/moTask agent
More frameworks agents
Listed but not yet reviewed in depth.
