langchain-middleware

You can inject custom logic before or after LangChain model calls by implementing middleware hooks. These hooks allow you to run validation, logging, or input shaping code at key execution points within your agent workflows.

The best way to intercept and modify LangChain tool calls is by using the wrap_tool_call middleware method. This allows you to intercept requests and responses to route them through dynamic logic or enforce guardrails during agent execution.

Yes, you can chain multiple middleware components together into a single flexible workflow. This composability allows you to combine various class-based or decorator-based hooks for complex agent behavior.

Yes, LangChain middleware supports asynchronous operation and context. You can define your middleware using class-based structures or decorator-based hooks while leveraging async capabilities for diverse agent workflows.

You can enforce guardrails or dynamically switch models by designing custom middleware for LangChain agents. By wrapping model calls, you intercept requests to apply guardrails or route logic based on user context.