@langchain/mcp-adapters weekly downloads statistics

LangChain.js MCP Adapters

Give your LangChain agents access to tools from Model Context Protocol (MCP) servers. @langchain/mcp-adapters manages connections and converts MCP tools and results into LangChain formats, ready to use with createAgent or a custom LangGraph workflow.

• Connect your tools: discover tools from multiple local or remote servers. • Control tool execution: authenticate connections, customize arguments and results with hooks, and receive progress updates. • Choose what the model sees: use text and multimodal results as model input, or keep outputs in artifacts for your application to process. • Ask for user input: pause an agent when a modern MCP server requests a form response or a URL visit, then resume with the user's response.

Read the MCP guide for concepts, configuration, and advanced usage.

Install

Requires Node.js 20.10 or later. Install the adapter and its LangChain peers:

The adapter requires @langchain/core ^1.2.6 and @langchain/langgraph ^1.4.13. It includes the MCP SDK client; install the SDK separately only if your application imports it directly.

Quickstart: give an agent MCP tools

This example connects an agent to the public LangChain docs MCP server so it can look up documentation. You do not need to run a server or configure authentication for this MCP endpoint.

Install LangChain and the model integration used below:

Set OPENAI_API_KEY in your environment, then run:

Keep the adapter open while your agent uses its tools, then call close() when finished. Connections open as needed. listTools() returns executable LangChain tools, which you can also invoke directly without a model or use in a custom LangGraph workflow.

Tools that request user input

MCP tools can ask users to complete a form or visit a URL before continuing. For modern MCP servers, the adapter pauses the agent through a LangGraph interrupt so your application can collect a response and resume the run.

Configure a checkpointer for these workflows. Tools that do not request input can run without one. See the tools guide for handling requests and resuming execution.

Documentation and examples

| I want to… | Start here |

| Connect local or remote servers, manage connections, or select tools | Connections | | Authenticate with tokens or OAuth | Authentication | | Customize tool calls, handle results, or collect user input | Tools | | Modify tool arguments and results | Hooks example | | Work with multimodal content and artifacts | Content example | | Receive server messages and tool progress | Notifications example | | Run a local example without model credentials | Example setup and walkthroughs |

If you already manage an MCP SDK client, use loadMcpTools() to adapt its tools without handing connection management to MCPAdapter.

Upgrading from 1.x

Use MCPAdapter, { servers: { ... } }, and listTools() for new code. MultiServerMCPClient, mcpServers, and getTools() remain available as deprecated compatibility APIs.

Version 2 also changes tool-name defaults, configuration, connection behavior, and tool results. Review the migration guide before upgrading, including any approval rules that refer to tool names.

Acknowledgements

Big thanks to @vrknetha, @knacklabs for the initial implementation!

Contributing

Contributions are welcome! See the contributing guidelines.

License

MIT

Description from the npm listing.