Connect Cheerful to Claude and ChatGPT

It is a remote server, so there is no local process, no npx wrapper and no configuration file to keep in sync across machines. Sign in once per client. https://mcp.cheerful.ai/mcp/healthz answers without a token if you need to prove the endpoint is reachable before blaming your client.

The tool chips are the part worth watching. Every call the assistant makes is one of nine named tools, each marked read-only, and your client shows you which ones ran. Nothing in this exchange writes anything: the assistant read two campaigns and their performance, and the answer is assembled from what it found.

Ask in plain English. The server ships an instruction map that tells the assistant which tool owns which question, so you do not have to name tools yourself.

The first six are the ones worth trying first. They are questions about work already in progress, and answering them by hand means opening a campaign, a mailbox and a billing page in three tabs.

Arguments, response shapes and the traps for each one are on the tools page. Searches consume credits at the same rate as the web app; every other tool is free.

These are lifted from the server’s own instructions, which is to say they are what the connector tells the model before you type anything. They are here because a docs page that hides them is a docs page you find out was wrong later.

• A filtered or smart creator search takes 90–155 seconds, and a cold creator profile about 60. That is the normal shape, not a hang. Retrying over the top of a running call spends credits twice and does not make it faster.
• Smart search only binds a location for fifteen countries. A brief naming anywhere else has its location silently dropped and comes back global — use a keyword search with an explicit country name instead.
• Most creators have no public email. That is the common case, not a failure, and not a reason to search again — emails are enriched when a creator is added to a campaign.
• A total in a search result is the provider’s pre-filter pool, often tens of thousands. It is not what you received, and an assistant that reports it as "found 40,000 creators" is reading it wrong.
• Campaigns can belong to a teammate. Reads span your own work plus everything your teams share, and each campaign carries the team_id and owner_email that say whose it is.
• Senders belong to the campaign, not to you. A mailbox you can read is not necessarily one you may send from.
• A 404 and "not yours" are deliberately the same response, so a 404 is not proof something does not exist.
• A 402 means you are out of credits rather than that something broke.

The connector uses OAuth 2.1. Your client discovers where to sign you in from the server’s protected-resource metadata, sends you to Cheerful, and receives a token issued to that client for your account. We never see your assistant’s conversation, and your assistant never sees a Cheerful password.

A connector token carries exactly your own permissions. It reads your campaigns and the ones your teams share with you, the mailboxes you have been granted, and nothing else. Revoking the connector in your client ends its access; a Cheerful session that has been revoked stops working through the connector at the same moment it stops working in the app.

What we store about connector use is what we store about any use of your account. The privacy policy covers it, and it is worth reading before you connect a work account to an assistant.

There is one other MCP server in this category, and it exposes creator lookups: audience data, engagement rates, estimated costs, contact details. Useful, and genuinely the same thing our creator search does.

The difference is everything around it. Cheerful’s connector reads campaigns, senders, the inbox, replies, teams and credits — the state of outreach that is already running. That is not a data product; it is the operational record of work in flight, and it is what makes "which campaign is underperforming and who is it sending from" answerable in one turn instead of five.