MCP to LangChain Tools Conversion Utility / TypeScript

A simple, lightweight library intended to simplify the use of Model Context Protocol (MCP) server tools with LangChain.

Its simplicity and extra features, such as tools schema adjustments for LLM compatibility and tools invocation logging, make it a useful basis for your experiments and customizations. However, it only supports text results of tool calls and does not support MCP features other than Tools.

LangChain's official LangChain.js MCP Adapters library, which supports comprehensive integration with LangChain, has been released. You may want to consider using it if you don't have specific needs for this library.

Introduction

This package is intended to simplify the use of Model Context Protocol (MCP) server tools with LangChain / TypeScript.

Model Context Protocol (MCP) is the de facto industry standard that dramatically expands the scope of LLMs by enabling the integration of external tools and resources, including DBs, Cloud Storages, GitHub, Docker, Slack, and more. There are [quite a few useful MCP servers already available. See MCP Server Listing on the Official Site.

This utility's goal is to make these numerous MCP servers easily accessible from LangChain.
It contains a utility function convertMcpToLangchainTools().
This async function handles parallel initialization of specified multiple MCP servers and converts their available tools into an array of LangChain-compatible tools. It also performs LLM provider-specific schema transformations to prevent schema compatibility issues

For detailed information on how to use this library, please refer to the following document: "Supercharging LangChain: Integrating 2000+ MCP with ReAct".
A Python equivalent of this utility is available here

Prerequisites

• Node.js 18+

Installation

npm i @h1deya/langchain-mcp-tools
Copy

API docs

Can be found here

Quick Start

A minimal but complete working usage example can be found in this example in the langchain-mcp-tools-ts-usage repo

convertMcpToLangchainTools() utility function accepts MCP server configurations that follow the same structure as Claude for Desktop, but only the contents of the mcpServers property, and is expressed as a JS Object, e.g.:

const mcpServers: McpServersConfig = {
  filesystem: {
    command: "npx",
    args: ["-y", "@modelcontextprotocol/server-filesystem", "."]
  },
  fetch: {
    command: "uvx",
    args: ["mcp-server-fetch"]
  },
  github: {
    type: "http",
    url: "https://api.githubcopilot.com/mcp/",
    headers: {
      "Authorization": `Bearer ${process.env.GITHUB_PERSONAL_ACCESS_TOKEN}`
    }
  },
};

const { tools, cleanup } = await convertMcpToLangchainTools(
  mcpServers, {
    // Perform provider-specific JSON schema transformations to prevent schema compatibility issues
    llmProvider: "google_gemini"
    // llmProvider: "openai"
    // llmProvider: "anthropic"
  }
);
Copy

This utility function initializes all specified MCP servers in parallel, and returns LangChain Tools (tools: StructuredTool[]) by gathering available MCP tools from the servers, and by wrapping them into LangChain tools.

When llmProvider option is specified, it performs LLM provider-specific schema transformations for MCP tools to prevent schema compatibility issues. Set this option when you enconter schema related warnings/errors while execution.
See below for details.

It also returns an async callback function (cleanup: McpServerCleanupFn) to be invoked to close all MCP server sessions when finished.

The returned tools can be used with LangChain, e.g.:

// import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
const llm = new ChatGoogleGenerativeAI({ model: "gemini-2.5-flash" });

// import { createReactAgent } from "@langchain/langgraph/prebuilt";
const agent = createReactAgent({
  llm,
  tools
});
Copy

A minimal but complete working usage example can be found in this example in the langchain-mcp-tools-ts-usage repo

For hands-on experimentation with MCP server integration, try this MCP Client CLI tool built with this library

Building from Source

See README_DEV.md for details.

MCP Protocol Support

This library supports MCP Protocol version 2025-03-26 and maintains backwards compatibility with version 2024-11-05. It follows the official MCP specification for transport selection and backwards compatibility.

Limitations

• Tool Return Types: Currently, only text results of tool calls are supported. The library uses LangChain's response_format: 'content' (the default), which only supports text strings. While MCP tools can return multiple content types (text, images, etc.), this library currently filters and uses only text content.
• MCP Features: Only MCP Tools are supported. Other MCP features like Resources, Prompts, and Sampling are not implemented.

Notes:

• LLM Compatibility and Schema Transformations: The library can perform schema transformations for LLM compatibility. See below for details.
• Passing PATH Env Variable: The library automatically adds the PATH environment variable to stdio server configrations if not explicitly provided to ensure servers can find required executables.

Features

stderr Redirection for Local MCP Server

A new key "stderr" has been introduced to specify a file descriptor to which local (stdio) MCP server's stderr is redirected.
The key name stderr is derived from TypeScript SDK's StdioServerParameters.

    const logPath = `mcp-server-${serverName}.log`;
    const logFd = fs.openSync(logPath, "w");
    mcpServers[serverName].stderr = logFd;
Copy

A usage example can be found here

Working Directory Configuration for Local MCP Servers

The working directory that is used when spawning a local (stdio) MCP server can be specified with the "cwd" key as follows:

    "local-server-name": {
      command: "...",
      args: [...],
      cwd: "/working/directory"  // the working dir to be use by the server
    },
Copy

The key name cwd is derived from TypeScript SDK's StdioServerParameters.

Transport Selection Priority

The library selects transports using the following priority order:

1. Explicit transport/type field (must match URL protocol if URL provided)
2. URL protocol auto-detection (http/https → StreamableHTTP → SSE, ws/wss → WebSocket)
3. Command presence → Stdio transport
4. Error if none of the above match

This ensures predictable behavior while allowing flexibility for different deployment scenarios.

Remote MCP Server Support

mcp_servers configuration for Streamable HTTP, SSE and Websocket servers are as follows:

    // Auto-detection: tries Streamable HTTP first, falls back to SSE on 4xx errors
    "auto-detect-server": {
        url: `http://${server_host}:${server_port}/...`
    },

    // Explicit Streamable HTTP
    "streamable-http-server": {
        url: `http://${server_host}:${server_port}/...`,
        transport: "streamable_http"
        // type: "http"  // VSCode-style config also works instead of the above
    },

    // Explicit SSE
    "sse-server-name": {
        url: `http://${sse_server_host}:${sse_server_port}/...`,
        transport: "sse"  // or `type: "sse"`
    },

    // WebSocket
    "ws-server-name": {
        url: `ws://${ws_server_host}:${ws_server_port}/...`
        // optionally `transport: "ws"` or `type: "ws"`
    },
Copy

For the convenience of adding authorization headers, the following shorthand expression is supported. This header configuration will be overridden if either streamableHTTPOptions or sseOptions is specified (details below).

    github: {
      // To avoid auto protocol fallback, specify the protocol explicitly when using authentication
      type: "http",  // or `transport: "http",`
      url: "https://api.githubcopilot.com/mcp/",
      headers: {
        "Authorization": `Bearer ${process.env.GITHUB_PERSONAL_ACCESS_TOKEN}`
      }
    },
Copy

NOTE: When accessing the GitHub MCP server, GitHub PAT (Personal Access Token) alone is not enough; your GitHub account must have an active Copilot subscription or be assigned a Copilot license through your organization.

Auto-detection behavior (default):

• For HTTP/HTTPS URLs without explicit transport, the library follows MCP specification recommendations
• First attempts Streamable HTTP transport
• If Streamable HTTP fails with a 4xx error, automatically falls back to SSE transport
• Non-4xx errors (network issues, etc.) are re-thrown without fallback

Explicit transport selection:

• Set transport: "streamable_http" (or VSCode-style config type: "http") to force Streamable HTTP (no fallback)
• Set transport: "sse" to force SSE transport
• WebSocket URLs (ws:// or wss://) always use WebSocket transport

Streamable HTTP is the modern MCP transport that replaces the older HTTP+SSE transport. According to the official MCP documentation: "SSE as a standalone transport is deprecated as of protocol version 2025-03-26. It has been replaced by Streamable HTTP, which incorporates SSE as an optional streaming mechanism."

Authentication Support for Streamable HTTP Connections

The library supports OAuth 2.1 authentication for Streamable HTTP connections:

import { OAuthClientProvider } from '@modelcontextprotocol/sdk/client/auth.js';

// Implement your own OAuth client provider
class MyOAuthProvider implements OAuthClientProvider {
  // Implementation details...
}

const mcpServers = {
  "secure-streamable-server": {
    url: "https://secure-mcp-server.example.com/mcp",
    // To avoid auto protocol fallback, specify the protocol explicitly when using authentication
    transport: "streamable_http",  // or `type: "http",`
    streamableHTTPOptions: {
      // Provide an OAuth client provider
      authProvider: new MyOAuthProvider(),
      
      // Optionally customize HTTP requests
      requestInit: {
        headers: {
          'X-Custom-Header': 'custom-value'
        }
      },
      
      // Optionally configure reconnection behavior
      reconnectionOptions: {
        maxReconnectAttempts: 5,
        reconnectDelay: 1000
      }
    }
  }
};
Copy

Test implementations are provided:

• Streamable HTTP Authentication Tests:
  • MCP client uses this library: streamable-http-auth-test-client.ts
  • Test MCP Server: streamable-http-auth-test-server.ts

Change Log

Can be found here

Appendix

Troubleshooting

1. Enable debug logging: Set logLevel: "debug" to see detailed connection and execution logs
2. Check server stderr: For stdio MCP servers, use stderr redirection to capture server error output
3. Test explicit transports: Try forcing specific transport types to isolate auto-detection issues
4. Verify server independently: Test the MCP server with other clients (e.g., MCP Inspector)

LLM Provider Schema Compatibility

Different LLM providers have incompatible JSON Schema requirements for function calling:

• OpenAI requires: Optional fields must be nullable (.optional() + .nullable()) for function calling (based on Structured Outputs API requirements, strict enforcement coming in future SDK versions)"
• Google Gemini API: Rejects nullable fields and $defs references, requires strict OpenAPI 3.0 subset compliance
• Anthropic Claude: Very relaxed schema requirements with no documented restrictions

Note: Google Vertex AI provides OpenAI-compatible endpoints that support nullable fields.

Real-World Impact

This creates challenges for developers trying to create universal schemas across providers.

Many MCP servers generate schemas that don't satisfy all providers' requirements.
For example, the official Notion MCP server @notionhq/notion-mcp-server (as of Jul 2, 2025) produces:

OpenAI Warnings:

Zod field at `#/definitions/API-get-users/properties/start_cursor` uses `.optional()` without `.nullable()` which is not supported by the API. See: https://platform.openai.com/docs/guides/structured-outputs?api-mode=responses#all-fields-must-be-required
... followed by many more
Copy

Gemini Errors:

GoogleGenerativeAIFetchError: [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent: [400 Bad Request] * GenerateContentRequest.tools[0].function_declarations[0].parameters.properties[children].items.properties[paragraph].properties[rich_text].items.properties[mention].any_of[0].required: only allowed for OBJECT type
... followed by many more
Copy

Solution

The new option, llmProvider, has been introduced for performing provider-specific JSON schema transformations:

const { tools, cleanup } = await convertMcpToLangchainTools(
  mcpServers, {
    llmProvider: "openai"        // Makes optional fields nullable
    // llmProvider: "google_gemini" // Applies Gemini's strict validation rules  
    // llmProvider: "anthropic"     // No transformations needed
  }
);
Copy

Features:

• Generates INFO-level logs when transformations are applied
• Helps users identify potential MCP server schema improvements
• Falls back to original schema when no llmProvider is specified

Provider-Specific Transformations

Provider	Transformations Applied
openai	Makes optional fields nullable, handles union types
google_gemini	Filters invalid required fields, fixes anyOf variants, removes unsupported features
anthropic	Accepts schemas as-is, but handles them efficiently

For other providers, try without specifying the option:

const { tools, cleanup } = await convertMcpToLangchainTools(
  mcpServers
);
Copy

References

• OpenAI Function Calling
• Gemini API Schema Requirements
• Anthropic Tool Use

Resource Management

The returned cleanup function properly handles resource cleanup:

• Closes all MCP server connections concurrently
• Logs any cleanup failures
• Continues cleanup of remaining servers even if some fail
• Should always be called when done using the tools

const { tools, cleanup } = await convertMcpToLangchainTools(mcpServers);

try {
  // Use tools with your LLM
} finally {
  // Always cleanup, even if errors occur
  await cleanup();
}
Copy

Debugging and Logging

The library provides configurable logging to help debug connection and tool execution issues:

// Configure log level
const { tools, cleanup } = await convertMcpToLangchainTools(
  mcpServers, 
  { logLevel: "debug" }
);

// Use custom logger
class MyLogger implements McpToolsLogger {
  debug(...args: unknown[]) { console.log("[DEBUG]", ...args); }
  info(...args: unknown[]) { console.log("[INFO]", ...args); }
  warn(...args: unknown[]) { console.warn("[WARN]", ...args); }
  error(...args: unknown[]) { console.error("[ERROR]", ...args); }
}

const { tools, cleanup } = await convertMcpToLangchainTools(
  mcpServers,
  { logger: new MyLogger() }
);
Copy

Available log levels: "fatal" | "error" | "warn" | "info" | "debug" | "trace"

For Developers

See README_DEV.md for more information about development and testing.