LangChain is an open-source Python and TypeScript framework for building LLM agents with a standard interface to hundreds of model and tool integrations. It supports tool calling, MCP servers, memory, structured output and tracing via LangSmith.
@langchain/typesafe 0.1.0 is a breaking release: pass questions to invoke, read usage_metadata, and import the middleware from langchain.
langchain 1.6.5 and @langchain/core 1.4.0 add decision models and two new middleware, and both now require Zod 4.6.6 or newer.
langchain 1.5.0 and core 1.8.0 need Python 3.11, add decision models, and the router and auto mode middleware now need classifier=.
LangChain.js now marks model refusals with a refusal block, and output parsers throw ModelRefusalError instead of a generic parse error.
Ask typed yes/no, choice and score questions with OpenAI's Decisions API from LangChain.js, plus model-routing and tool-risk middleware.
Tool-call arguments now stream as tool-call-delta events. Old clients break, so upgrade them before the server.
Wirestack