Libraries you import to build agents. Not chat UIs, not MCP servers, not SKILL.md packs. One strong option per style: chains, graphs, multi-agent, typed Python, first-party SDKs.
English-speaking builders choosing how to orchestrate models and tools. If you want something you can docker-compose, use the self-hosted list instead.
Icon, name, one sentence. Open a card for stars and the repo detail page.
The default Python/JS toolkit for chaining models, tools, and retrieval.
Still the comparison point for “how do I wire an LLM to tools” in English search.
Stateful, graph-based agent workflows on top of the LangChain ecosystem.
The production-shaped successor when a linear chain is not enough.
Role-based multi-agent crews for research, writing, and ops-style jobs.
The most-searched multi-agent productized framework after LangChain.
Microsoft AutoGen for multi-agent conversations and tool-using workflows.
Enterprise-familiar name; useful as the Microsoft-stack alternative on the list.
Pydantic-first agents: structured outputs, tools, and type-safe Python.
Fills the “I want agents that look like normal Python” slot.
OpenAI’s official Python agents SDK for tool calling and handoffs.
First-party SDK for teams already standardized on OpenAI APIs.
Small Hugging Face agents that write and run code to finish a task.
Lightweight alternative to heavy orchestration graphs.
Google Agent Development Kit for building agents on Gemini and Vertex.
The Google-cloud counterpart so the list is not only OpenAI/LangChain.
Agno (formerly Phidata) for fast agents with tools, memory, and teams.
High-velocity Python agent runtime that shows up in “faster than LangChain” searches.
LlamaIndex for data-connected agents and retrieval over private documents.
The RAG-native framework when the job is “talk to my files,” not generic chat.
Self-hosted AI is products you run (Open WebUI, Dify, Ollama). This page is frameworks you import to build your own agent. Different search intent, different shortlist.
Not automatically. Use LangGraph or LlamaIndex if you already know the job is stateful workflows or RAG. Use PydanticAI or the OpenAI Agents SDK if you want less framework and more typed Python.
MCP servers give agents tools. Skills give agents procedures. Frameworks glue models, tools, and state. They are three separate pages on purpose.
Weekly. Renames (for example Phidata → Agno) are updated on the next pass.