Langchain Mcp Multi Server

Follow on X WhatsApp Channel License

Python LangChain LangGraph MCP NVIDIA NIM uv

An educational example showing how a single LangChain agent can call tools from multiple MCP servers at once — one over stdio, one over SSE — powered by a free NVIDIA NIM model (nvidia/nemotron-3-ultra-550b-a55b) instead of OpenAI.

flowchart LR
    U([User question]) --> A["LangChain agent<br/>(create_agent + ChatNVIDIA)"]
    A -->|stdio · subprocess| M["Math MCP server<br/>add · multiply"]
    A -->|SSE · http://localhost:8000/sse| W["Weather MCP server<br/>get_weather"]
    A --> R([Final answer])

---

📖 The two MCP transports, explained

MCP separates what a server offers (tools, prompts, resources) from how you talk to it (the transport). This repo demonstrates both classic transports:

	🖥️ stdio	🌐 SSE (Server-Sent Events)
How it works	The client spawns the server as a subprocess and speaks JSON-RPC over stdin/stdout	The server is a standalone web service; clients connect over HTTP
Who starts the server	The client, automatically	You, manually (or a process manager)
Clients per server	Exactly one	Many simultaneous clients
Best for	Local tools, CLI integrations, Claude Desktop	Shared/remote services, microservices
In this repo	servers/math_server.py	servers/weather_server.py

ℹ️ Newer MCP versions introduce Streamable HTTP as the successor to SSE for remote servers. SSE is still widely deployed and is what this tutorial (and most courses) teach — the client-side code barely changes.

🗂️ Project structure

.
├── servers/
│   ├── math_server.py      # MCP server #1 — stdio transport (add, multiply)
│   └── weather_server.py   # MCP server #2 — SSE transport on port 8000 (get_weather, mocked)
├── main.py                 # Example 1: agent + ONE server (stdio)
├── langchain_client.py     # Example 2: agent + MULTIPLE servers (stdio + SSE)
├── assets/                 # Screenshots
├── .env.example            # Template for your NVIDIA API key
└── pyproject.toml          # Dependencies (managed with uv)

🔑 Get a free NVIDIA API key

This project uses NVIDIA NIM hosted endpoints — free to try, no credit card:

1. Go to build.nvidia.com and sign in
2. Open any model page and click Get API Key
3. Copy the key (starts with nvapi-)
4. Create your .env file:

cp .env.example .env      # then paste your key inside

NVIDIA_API_KEY=nvapi-xxxxxxxxxxxxxxxxxxxxxxxx
# Optional — any NIM model that supports tool calling:
# NVIDIA_MODEL=nvidia/nemotron-3-ultra-550b-a55b

⚠️ The model must support tool calling for the agent to work. nvidia/nemotron-3-ultra-550b-a55b does. Browse other models in the NIM API reference.

🚀 Setup

Prerequisites: Python 3.12+ and uv.

git clone <this-repo>
cd sse-mcp
uv sync

▶️ Example 1 — One agent, one stdio server

No server to start — the client spawns math_server.py itself:

uv run main.py

Expected output:

MCP session initialized
Loaded tools: ['add', 'multiply']
The expression 54 + 2 × 3 follows the order of operations ...
**Answer: 60**

What happened under the hood:

1. stdio_client(...) spawned the math server as a subprocess
2. load_mcp_tools(session) converted its MCP tools into LangChain tools
3. create_agent(llm, tools) built an agentic loop around the NVIDIA model
4. The model decided to call multiply(2, 3) then add(54, 6) — MCP carried each call to the server and the result back

▶️ Example 2 — One agent, multiple servers (stdio + SSE)

Terminal 1 — start the SSE weather server first:

uv run servers/weather_server.py

Wait for: Uvicorn running on http://localhost:8000

Terminal 2 — run the multi-server client:

uv run langchain_client.py

Expected output:

Loaded tools: ['add', 'multiply', 'get_weather']
The current weather in San Francisco is **14°C, foggy with a light breeze**.
And **2 + 2 = 4**.

The magic is MultiServerMCPClient: it aggregates tools from any number of servers behind different transports into one flat list — the agent never knows (or cares) where a tool lives:

client = MultiServerMCPClient(
    {
        "math":    {"command": "python", "args": [MATH_SERVER], "transport": "stdio"},
        "weather": {"url": "http://localhost:8000/sse", "transport": "sse"},
    }
)
tools = await client.get_tools()   # ['add', 'multiply', 'get_weather']

🔄 Differences from the original (OpenAI) course code

This example is adapted from a Udemy MCP course that uses ChatOpenAI. What changed and why:

Change	Why
ChatOpenAI() → ChatNVIDIA(model=..., timeout=300)	Use NVIDIA NIM's free endpoints; timeout=300 because free endpoints can queue requests beyond the default 60 s read timeout
Hardcoded absolute server path → Path(__file__).parent / "servers" / ...	Works wherever the repo is cloned, on any OS
langgraph.prebuilt.create_react_agent → langchain.agents.create_agent	The former is deprecated since LangGraph 1.0; create_agent is the modern replacement (same ReAct loop underneath)
"transport": "stdio" added explicitly	Required by current langchain-mcp-adapters

🛠️ Troubleshooting

Symptom	Fix
SocketTimeoutError: Timeout on reading data from socket	The NIM endpoint is queueing your request — the timeout=300 in the code covers most cases; retry, or switch NVIDIA_MODEL to a smaller model
ConnectionError on localhost:8000	The weather server isn't running — start it first (Example 2, Terminal 1)
[Errno 10048] ... port is in use	Another process holds port 8000 — stop it or change the port in weather_server.py (and the URL in langchain_client.py)
Garbled characters like 14�C in the terminal	Cosmetic Windows console encoding issue — run chcp 65001 or ignore it

🌍 Discover more MCP servers

You don't have to write every server yourself — there is a huge ecosystem you can plug into the same MultiServerMCPClient:

• awesome-mcp-servers — a curated GitHub list with hundreds of production-ready and community MCP servers (databases, browsers, GitHub, Slack, filesystems...), organized by category.
• Glama MCP directory — a searchable directory of 55,000+ MCP servers with filters by language, transport (remote/local), and category, plus its own inspector:

📚 Learn more

• Model Context Protocol — transports
• langchain-mcp-adapters — the bridge between MCP tools and LangChain
• LangGraph documentation
• NVIDIA NIM LLM APIs
• ChatNVIDIA integration docs

📄 License

MIT — see LICENSE.

---

Made with ❤️ for the MCP community. Follow @mcoding_off for more tutorials.