Connect your business to Claude, ChatGPT, Cursor and VS Code

Outcome: your own business answering inside the AI tools you already use — ask an assistant about your hours or prices and watch it fetch them from your profile instead of guessing.

You need: a published profile. No coding, but you will paste a line of configuration into an application.

Why point an assistant at yourself

It proves the endpoint works end to end from the outside, and it shows you exactly what an agent sees — the fastest way to notice a service you never listed.

Step 1 — Publish first, then read the badge

Settings → AI Agents → Connect an AI assistant. Three badge states, each sending you to a different control: Live, the endpoint answers; Off, you are not published yet (the config stays copyable, it just will not get an answer); Paused, you are published but direct agent connections are switched off a few rows above.

Step 2 — Open the setup and choose your client

Press Show setup. Five tabs, and they genuinely differ, because these clients disagree about how a remote MCP server is added. Nothing is typed by hand: each block is built from your own slug, so the address is always the live one.

Step 3 — Paste it where that client expects it

Claude Desktop — a settings dialog, not a config file: Settings → Connectors → Add custom connector, paste the address, save. (The mcpServers map in its config file is the mechanism for local servers; a remote address there connects nothing.)

ChatGPT — also a dialog: Settings → Connectors, turn on Developer mode under Advanced settings, then Connectors → Create, paste, save. Whether this appears depends on your ChatGPT plan — check OpenAI's own documentation, since their packaging changes independently of us.

Claude Code — one line in a terminal:

claude mcp add --transport http your-slug https://coreloop.so/api/mcp/your-slug

Cursor — a url entry in the mcpServers map, in .cursor/mcp.json or in Cursor’s global MCP settings.

VS Code — .vscode/mcp.json, and a genuinely different shape: a servers map with an explicit "type": "http".

Step 4 — Ask something only your profile knows

Not “what is my business”, which any model will bluff. Ask your Saturday hours, the price of a named service, whether you take card. Six tools sit behind that address: your details, your services, your opening hours, your reviews, a comparison tool that appears once a single category holds two or more services, and the one that sends you a message.

Step 5 — Know the ceiling and the off switch

Hourly request ceiling per business, by plan: 100 on Free, 1,000 on Starter, 10,000 on Pro and Agency, 100,000 on Enterprise. Ordinary use never approaches it; it exists so one noisy caller cannot drown yours. The switch that turns direct agent connections off is in the same card; your llms.txt deliberately has none, and stays available while you are published.

Done when

An AI client answers a specific question about your business correctly, and the panel reads Live.

Worth knowing: an agent that does not know your slug finds you another way. coreloop.so/.well-known/mcp.json is CoreLoop’s own machine-readable card: it points a crawler at the directory, and carries the URL patterns for every business’s MCP endpoint, A2A endpoint, agent card and llms.txt. You configure none of it. More: the protocols overview and the directory.