Orchestrating AI Tools with LangGraph and MCP Server

Setup

pip install langchain-mcp-adapters

pip install langgraph "langchain[openai]"

MCP Server definitions

Two MCP servers are created according to the Model Context Protocol (MCP) specification.

StdIO‑based math server provides two tools: add(a, b) (returns the sum of two numbers) and multiply(a, b) (returns the product). The server runs as a subprocess of the LangGraph agent and communicates via a stdio pipe. The division tool divide(a, b) is intentionally omitted.

HTTP weather server provides a single tool get_weather(city) that returns the weather for the specified city. The server is exposed as a RESTful API and is invoked over HTTP.

Agent client

The client creates a MultiServerMCPClient() instance, which connects to both the math server (stdio) and the weather server (HTTP). Calling await client.get_tools() returns a list of the three discovered tool specifications. These tools are bound to a LangGraph ReAct agent via create_react_agent(), enabling the agent to invoke the tools during reasoning.

Interaction examples

Math expression : Query "(3 + 5) x 12" . The agent first calls add(3, 5) → 8, then calls multiply(8, 12) → 96. The final answer is 96.

Division request : Query "what's the result divided by 6?" . Because the divide tool is missing, the agent mistakenly reuses the add tool with arguments 96 and -90, producing 6. The author notes that adding a divide tool would correct this behavior.

Weather query : Query "what is the weather in Beijing?" . The agent invokes the HTTP get_weather("Beijing") tool, receives the weather data, and stores the full dialogue context.

LangGraph core concepts

Node : A Python function that receives the current dialogue state and returns an updated state. Nodes encapsulate single responsibilities such as prompting the LLM, invoking a tool, or performing logical checks.

Edge : Connects nodes and defines execution order. Edges support conditional branching, loops, and termination, allowing the graph to model complex agent workflows.

References

GitHub repository: https://github.com/langchain-ai/langchain-mcp-adapters

Test article: https://zhuanlan.zhihu.com/p/1893223123081482311