KeyError: 'transport' from MultiServerMCPClient.get_tools()

KeyError: 'transport' from MultiServerMCPClient.get_tools()

TL;DR: your server config is wrapped in an extra "mcpServers" key that MultiServerMCPClient does not expect. Pass the servers dict directly: MultiServerMCPClient({"airbnb": {"transport": "stdio", ...}}). The client iterates the dict you give it, so one extra nesting level breaks the "transport" lookup.

KeyError: 'transport'
# raised by await client.get_tools() when the config dict has the wrong shape.

When this applies

• You use MultiServerMCPClient from langchain_mcp_adapters to load MCP tools into a LangGraph agent.
• get_tools() raises KeyError: 'transport' immediately, before any server starts.
• Your config looks like {"mcpServers": {"airbnb": {...}}} (copied from a blog post or the FastMCP multi-server docs).

When it does not

• Timeouts or connection refused mean the server config is right but the server is not reachable. Different fix.
• get_tools() returning an empty list means the server started but exposed no tools.

Fix it

1. Drop the wrapper key

# wrong: extra "mcpServers" nesting
client = MultiServerMCPClient(connections={"mcpServers": {"airbnb": {...}}})

# right: servers dict directly
from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "airbnb": {
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "@openbnb/mcp-server-airbnb"],
    }
})
tools = await client.get_tools()

Expected: get_tools() returns the tool list instead of raising KeyError.

2. Every server entry needs a transport

# stdio servers
"my_server": {"transport": "stdio", "command": "npx", "args": [...]}

# HTTP servers
"other": {"transport": "streamable_http", "url": "YOUR_HOST:4001/mcp"}

Expected: the client knows how to reach each server. A missing "transport" key on ANY entry raises the same KeyError, so check all of them.

3. Then wire the tools into your agent

from langgraph.prebuilt import create_react_agent

agent = create_react_agent(model, tools)

Expected: the agent lists the MCP tools and can call them.

Why it happens

MultiServerMCPClient expects {server_name: server_config} and reads server_config["transport"] for each entry. When you wrap everything in "mcpServers", it iterates one entry named "mcpServers" whose value is the whole servers dict, and that dict has no "transport" key. The blog-post shape comes from a different client (FastMCP's multi-server config), not this one.

Edge cases

• connections= vs positional: both spellings appear in examples. The SHAPE of the dict matters, not the kwarg name.
• npx-based servers need node installed where the agent runs. In Docker, that means node in the image, not just on your laptop.
• On Windows, stdio servers often need shell: True or a .cmd shim for npx.

Compatibility

langchain-mcp-adapters (Python), used with langgraph's create_react_agent / ToolNode.