This simple Model Context Protocol (MCP) client demonstrates the use of MCP server tools by LangChain ReAct Agent.

It leverages a utility function convert_mcp_to_langchain_tools() from langchain_mcp_tools.
This function handles parallel initialization of specified multiple MCP servers and converts their available tools into a list of LangChain-compatible tools (List[BaseTool]).

LLMs from Anthropic, OpenAI and Groq are currently supported.

A typescript version of this MCP client is available here

• Python 3.11+
• [optional] uv (uvx) installed to run Python package-based MCP servers
• [optional] npm 7+ (npx) to run Node.js package-based MCP servers
• API keys from Anthropic, OpenAI, and/or Groq as needed

1. Install dependencies:

   make install

2. Setup API keys:

   cp .env.template .env

   • Update .env as needed.
   • .gitignore is configured to ignore .env to prevent accidental commits of the credentials.

3. Configure LLM and MCP Servers settings llm_mcp_config.json5 as needed.

   • The configuration file format for MCP servers follows the same structure as Claude for Desktop, with one difference: the key name mcpServers has been changed to mcp_servers to follow the snake_case convention commonly used in JSON configuration files.
   • The file format is JSON5, where comments and trailing commas are allowed.
   • The format is further extended to replace ${...} notations with the values of corresponding environment variables.
   • Keep all the credentials and private info in the .env file and refer to them with ${...} notation as needed.

Run the app:

make start

It takes a while on the first run.

Run in verbose mode:

make start-v

See commandline options:

make start-h

At the prompt, you can simply press Enter to use example queries that perform MCP server tool invocations.

Example queries can be configured in llm_mcp_config.json5