Package Layout

LangChain's packages form a dependency tree. Understanding the direction of that tree tells you which package to import an abstraction from, and which packages are safe to depend on for the long term.

Every arrow points at langchain-core. That is the stability contract:

Tier	Stability	Import abstractions from it?
langchain-core	Most stable — breaking changes are rare and loudly flagged	Yes. Runnable, BaseMessage, BaseChatModel all live here
langchain	Stable within a major version	For chains, agents, and orchestration helpers
Partner packages (langchain-openai, langchain-anthropic, ...)	Stable, maintained by the provider or LangChain team	For provider-specific model/embedding classes
langchain-community	No stability guarantee — anyone can contribute an integration here	Only when no partner package exists for what you need

tip

When you're writing type hints or building an abstraction that should work across providers, import from langchain-core (e.g. from langchain_core.language_models import BaseChatModel), not from a specific partner package. That is the one piece of the tree every provider package already depends on.

See also

• Installation — installing the packages described here.
• Keys and config — configuring the provider packages once installed.
• Runnables and LCEL — the Runnable abstraction that langchain-core defines.