What it is
LangChain is an open-source framework for building applications on top of large language models. The Python package (pip install langchain, MIT-licensed) lets a developer compose "chains" of prompts, models, retrievers and tools, and its create_agent API builds agents as, in the maintainers' words, "Model + Harness". A JavaScript/TypeScript version, LangChain.js, mirrors the same concepts.
Interface and dependencies
Most work happens through per-provider packages such as langchain-openai or langchain-anthropic, which give a consistent interface over model providers including OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI, Groq and Ollama. The documentation lists 1,000+ integrations spanning chat and embedding models, vector stores (Pinecone, Chroma, Qdrant, Weaviate, Milvus, Elasticsearch), document loaders, and tools such as Tavily search or Neo4j. The Python package requires Python 3.10 or later. Agent orchestration underneath is handled by LangGraph, a separate lower-level library from the same organisation.
What you build with it
A developer typically uses it to wire a chat model to a retriever over a vector store for retrieval-augmented generation, or to give a model tool-calling access to external APIs and let it run as an agent. Tracing and evaluation of what the chains and agents do is handled by a separate paid product, LangSmith, not by the open-source library itself.
Limits
LangChain's own current documentation frames the library primarily around building agents ("Agent = Model + Harness") rather than the general chain-composition tool it started as, so some older chain-based patterns are being superseded. Production observability, monitoring and evaluation require the separate commercial LangSmith platform.
