LangChain is an open-source framework for building applications around language models — the connective layer between a model and everything it needs to be useful: data sources, tools, memory, other models. It rose with the first wave of LLM applications by solving the unglamorous problem every team hit at once: the model is an API call, but the application is retrieval, parsing, tool invocation, and orchestration, and writing that plumbing bespoke for every project is waste. LangChain packaged the plumbing — and its integration catalogue, hundreds of connectors deep, remains the practical reason teams adopt it.
The framework's vocabulary maps cleanly to what you already know. Chains compose steps — prompt, model, parser, retrieval — into pipelines. Tools wrap capabilities the model may invoke, the same concept the Model Context Protocol now standardises across vendors. Memory carries state between turns. Agents, in LangChain's usage, are chains where the model chooses the next step — the agent loop in its toolkit form. For retrieval-heavy applications, the components assemble into the standard RAG architecture with unusually little code.
Where LangChain ends and LangGraph begins
The most common point of confusion is LangChain's relationship to LangGraph, its sibling from the same team. The short version: LangChain is the toolkit — components and integrations, control flow mostly linear and abstracted; LangGraph is the agent runtime — workflows as explicit graphs with first-class state, persistence, and human-approval stops. The full comparison draws the line in detail, but the practical reading is that the team's own centre of gravity for *agents* moved to LangGraph, with LangChain supplying the parts. If your system is a pipeline, LangChain alone is often enough; if it is an agent with consequences, you will probably end up drawing the graph.
What choosing LangChain settles — and what it cannot
A framework choice settles your development ergonomics: how fast integrations arrive, how much plumbing you skip, what your team debugs at 2am. It settles none of the readiness questions. A LangChain agent still needs its own identity and scoped credentials, evaluation before autonomy, a registry entry, and governance with names attached — the framework ships none of that, and no framework does. The frameworks landscape page makes the general argument: the framework is the most replaceable layer in an agentic system, so keep the durable parts — tools, evaluations, audit requirements — portable, and the cost of being wrong stays small.