Lovie MCP›OpenAI API

API

OpenAI API

The Responses API takes a remote MCP server as a tool. It is not an interactive OAuth client, so unlike every other client on this page you supply the access token yourself.

The only value you need

• Time 5 minutes
• Surface Responses API
• Auth token you supply
• Tool type mcp

The tool block

A remote MCP server is one entry in tools. The model discovers the tool list itself; you only name the server.

{
  "type": "mcp",
  "server_label": "lovie",
  "server_url": "https://mcp.lovie.co/mcp",
  "server_description": "Banking, cap table, accounting and US company formation.",
  "require_approval": "always",
  "authorization": "<lovie access token>"
}

Field	Meaning
server_url	The endpoint. A custom server always uses this; connector_id is only for OpenAI’s own hosted connectors.
server_label	How the model refers to the server in its output.
require_approval	"always" pauses for your confirmation before each call. Use "never" only once you trust the flow.
allowed_tools	Optional allowlist. Narrowing the surface improves tool choice and blast radius.
authorization	The OAuth access token. The API does not store it — send it on every request.

Getting a token

1. Install the CLI: brew install lovieco/tap/lovie — or see https://cli.lovie.co.
2. Run lovie auth login and complete the sign-in in your browser.
3. Read the access token out of the CLI’s credential store:

# macOS
jq -r '.credentials[] | select(.endpoint=="https://mcp.lovie.co/mcp") | .accessToken' \
  ~/Library/Application\ Support/lovie/credentials.json

# Linux
jq -r '.credentials[] | select(.endpoint=="https://mcp.lovie.co/mcp") | .accessToken' \
  ~/.config/lovie/credentials.json

Tokens expire. lovie auth login refreshes them, so read the value at request time rather than baking it into an image or a committed .env.

Python

import os
from openai import OpenAI

client = OpenAI()

resp = client.responses.create(
    model="gpt-5.6",
    tools=[{
        "type": "mcp",
        "server_label": "lovie",
        "server_url": "https://mcp.lovie.co/mcp",
        "server_description": "Banking, cap table, accounting and US company formation.",
        "require_approval": "always",
        "authorization": os.environ["LOVIE_ACCESS_TOKEN"],
    }],
    input="What is my cap table?",
)

print(resp.output_text)

TypeScript

import OpenAI from "openai";

const client = new OpenAI();

const resp = await client.responses.create({
  model: "gpt-5.6",
  tools: [{
    type: "mcp",
    server_label: "lovie",
    server_url: "https://mcp.lovie.co/mcp",
    server_description: "Banking, cap table, accounting and US company formation.",
    require_approval: "always",
    authorization: process.env.LOVIE_ACCESS_TOKEN,
  }],
  input: "What is my cap table?",
});

console.log(resp.output_text);

Without a token

Omit authorization and the connection still works: discovery is anonymous and the open tools answer. Anything that touches a company returns an authorization challenge instead of data, which is a reasonable way to explore the tool surface before wiring credentials.

If something goes wrong

Symptom	Cause
Tools listed, every call refused	No token, an expired one, or it was dropped — the API keeps nothing, so it must be on every request.
The model picks the wrong tool	Set allowed_tools to the handful the task needs.
Approvals on every step	Expected under "always". Move to "never" only for flows you have watched end to end.
Empty tool list	server_url is missing the /mcp path.

Somewhere else