Connect a LangGraph agent with LangChain MCP

Prerequisites

From setup to the first capture

• Python with langchain[mcp]>=1.4.0
• GRABBIT_API_KEY set by a secret manager
• LANGCHAIN_MODEL plus its provider package and credential, such as langchain-anthropic and ANTHROPIC_API_KEY

# pip install "langchain[mcp]==1.4.0" fastmcp langchain-anthropic
import asyncio
import os
from fastmcp.client import Client
from langchain.agents import create_agent
from langchain.mcp import MCPAdapter

async def run():
    client = Client(
        "https://mcp.grabbit.live",
        auth=os.environ["GRABBIT_API_KEY"],
    )
    async with MCPAdapter(client) as adapter:
        tools = await adapter.list_tools()
        agent = create_agent(os.environ["LANGCHAIN_MODEL"], tools)
        return await agent.ainvoke({
            "messages": [{
                "role": "user",
                "content": "Capture https://example.com and return the screenshot URL with a short visual QA report.",
            }]
        })

print(asyncio.run(run()))

FastMCP accepts a bearer-token string through Client(auth=...). It also supports auth="oauth" for a full OAuth 2.1 flow.

A test key returns a placeholder for wiring. The visual check below requires a live key and available credits. Call get_usage before a run when the agent needs to avoid an HTTP 402 interruption.

Troubleshooting

Flat pricing, built for agents that loop

When an agent captures a page on every step, per-call metering turns into anxiety. Grabbit is one flat rate: $50 per year for 25,000 captures ($0.002 each). Annual subscription credits expire at term end; separately purchased top-ups do not expire while your account remains active. Test-environment keys render placeholders for integration checks. A real page capture needs a live key and available credits.