Treeunfe API home pagelight logo
  1. AI and MCP
  2. Your own agent with the Claude API

With the Claude API MCP connector, your backend sends a message and Claude calls the Treeunfe tools on its own. You don’t write the tool loop or run an MCP client: Anthropic makes the connection to the server.

​
Before you start

  • A Claude API key (ANTHROPIC_API_KEY), created in the Anthropic Console.
  • The Treeunfe organization API Key (TREEUNFE_API_KEY), preferably a homologation key.

​
How it works

The request carries two parts that go together:

ParameterWhat it does
mcp_serversSays where the server is (url), names it (name) and sends the key in authorization_token. The key is sent as Authorization: Bearer.
toolsAn mcp_toolset entry with the same server name, which makes its tools available to Claude.

The connector is in beta and needs the anthropic-beta: mcp-client-2025-11-20 header.

​
Example

  • Python

  • TypeScript

pip install anthropic

import os
import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

response = client.beta.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    betas=["mcp-client-2025-11-20", "server-side-fallback-2026-07-01"],
    fallbacks="default",
    mcp_servers=[
        {
            "type": "url",
            "url": "https://mcp.treeunfe.com.br/mcp",
            "name": "treeunfe",
            "authorization_token": os.environ["TREEUNFE_API_KEY"],
        }
    ],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "treeunfe"}],
    messages=[
        {
            "role": "user",
            "content": "Look up the last 5 NF-e issued this month and summarize the amount and status of each.",
        }
    ],
)

if response.stop_reason == "refusal":
    print("The request was declined.")
else:
    for block in response.content:
        if block.type == "text":
            print(block.text)

fallbacks: "default" makes the API try another Claude model if the main model declines the request. You can remove this field and the server-side-fallback-2026-07-01 beta if you don’t want that behavior.

​
Limit the tools

For a read-only agent, turn every tool off by default and turn on only the read tools with default_config and configs:

tools=[
    {
        "type": "mcp_toolset",
        "mcp_server_name": "treeunfe",
        "default_config": {"enabled": False},
        "configs": {
            "READ_TOOL_NAME": {"enabled": True},
        },
    }
]

Replace READ_TOOL_NAME with the name shown in the server’s tool list. This adds to the key’s scope: even if a tool is enabled, a key without nfe:emit can’t issue invoices.

With the MCP connector, Claude calls tools without asking for approval. To issue in production, add a step where a person confirms the invoice data before it’s sent, or give the agent a key with only nfe:read and issue through the REST API after confirmation.