LangChain

LangChain is an open-source framework for building applications on top of large language models. It provides the plumbing most LLM apps end up needing - prompt templates, connectors to model providers, document loaders, retrieval over vector stores, tool calling, memory and multi-step agents - as composable building blocks in Python and JavaScript. Its rise tracked the early wave of LLM development, when every team was writing the same glue code to chain a model call to a search step to another model call. LangChain gave that pattern a shared vocabulary and a huge library of integrations, which is still its main strength: if a model, database or API exists, there is probably a LangChain connector for it. The company behind it has since built around the framework. LangGraph handles stateful, controllable agent workflows modelled as graphs, which addresses the most common complaint about early agent code - that it was hard to debug and hard to keep on the rails. LangSmith is a hosted platform for tracing, evaluating and monitoring LLM applications in production, and it works whether or not you use the framework itself. The criticism has been consistent: abstraction overhead. For simple applications, wrapping a single API call in several layers of LangChain classes can make code harder to read than calling the provider SDK directly, and the API surface changed quickly in the framework's early years. Many experienced teams use LangChain selectively - for integrations and LangSmith observability - rather than for everything. The alternatives depend on the job. LlamaIndex is more focused on retrieval and data ingestion for question-answering over documents. Haystack offers a similar pipeline approach with a search heritage. Provider SDKs from OpenAI and Anthropic, plus lightweight agent libraries, cover many cases with less machinery. Hugging Face is where you go for open models themselves. Choose LangChain when you need broad integrations and production observability for a non-trivial LLM application, and keep the abstractions only where they earn their place.