The Model Context Protocol (MCP) has seen significant updates within LangChain. MCP tool calls from ChatGPT users have increased by 98x across 2026, with a doubling in August alone. This update focuses on enabling a stateless core, a highly requested feature for developers seeking greater server reliability and scalability.
The core change is the transition to a stateless protocol. Previously, every MCP call required a session, pinning clients to specific server instances. The new specification eliminates this dependency, allowing for redeployments without disrupting live sessions. This shift unlocks two key capabilities: caching and elicitation. Caching allows servers to define tool list freshness, reducing redundant fetches, while elicitation enables tools to pause and request confirmation from the caller without holding connections open.
FastMCP, the underlying transport layer, provides abstractions for connections, authentication, caching, and protocol negotiation. The client surface is directly available, simplifying agent integration. The update introduces two distinct "eras" of the protocol, handled per connection by FastMCP, ensuring compatibility with both upgraded and legacy servers. The client can negotiate the newest era or fall back to the handshake for servers that haven't upgraded.
To facilitate adoption, the MCP support has been integrated into the main LangChain package. The MultiServerMCPClient collapses into a single MCPAdapter class, and a migration guide is available. Agents can now utilize tools like create_deep_agent with the mcp extra. The client supports various authentication methods, including bearer tokens, OAuth 2.1, and machine-to-machine credentials, and offers flexible transport options like streamable HTTP and in-memory connections. Source: https://www.langchain.com/blog/mcp-in-langchain-stateless-protocol-elicitation-and-more
