LangGraph shipped 1.0 as the production standard for agent orchestration, and the fastest way to give its graphs real tools is Model Context Protocol. Instead of hand-wiring every integration into your Python code, you point a LangGraph node at an MCP server and its tools appear as native LangChain tools. Here's the clean path — vetted, minimal, and the same shape whether you run one agent or a swarm.
The bridge: langchain-mcp-adapters
The official glue is the langchain-mcp-adapters package. It converts each MCP tool's JSON schema into a langchain_core BaseTool that any LangGraph node can call, so servers behave exactly like tools you wrote by hand. You need Python 3.11 or newer — the adapters won't load on older runtimes.
pip install langgraph langchain-mcp-adapters
Wire up a client
MultiServerMCPClient connects to one or many servers and returns their tools in a single list. Point it at a local stdio server or a remote Streamable HTTP endpoint, then hand the tools to a prebuilt ReAct agent:
from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent
client = MultiServerMCPClient({
"github": {"transport": "streamable_http", "url": "https://api.githubcopilot.com/mcp/"},
})
tools = await client.get_tools()
agent = create_react_agent("anthropic:claude-opus-4-8", tools)
Multi-agent graphs
The real payoff is at scale. LangGraph gives you checkpointing, human-in-the-loop pauses and durable state; MCP gives every node in the graph a live, versioned, network-accessible toolbox. A supervisor can route work to specialist agents that each own a different server — one on GitHub, one on a database, one on search — without any of them sharing code. That separation is what keeps a large graph auditable and lets you swap a tool without touching agent logic.
Which servers to start with
Give a coding graph the GitHub server and the filesystem server; give a research graph Exa or Firecrawl. Because the adapter speaks standard MCP, anything in our directory drops in unchanged — curated, categorised and vetted so you're not wiring a random GitHub repo into a production graph.
Going further
Deciding stdio vs HTTP? See transports compared. Building the servers yourself starts with the Python MCP tutorial. For orchestration theory, read multi-agent orchestration patterns, and browse the ML engineer loadout. Keep third-party servers vetted with MCP security best practices.