What is the Model Context Protocol (MCP)?
The Model Context Protocol, or MCP, is an open standard that lets an AI assistant call tools and read data from another system. The system publishes a list of tools, and any MCP-capable assistant can use them with the user's permission.
Example
A manager in a chat assistant types "what is waiting for my approval?". The assistant calls a tool on the company's MCP server with her access token, gets two pending requests, and summarizes them. She says "approve the first". The assistant calls the answer tool, and the server checks she is allowed to.
The assistant never has her password. The token decides what it can do.
In Atta
Atta runs an MCP server at /api/mcp. A bearer token is created per member, stored hashed, and checked for membership on every call, so removing a member stops it working. Tokens are also what Slack and WhatsApp links point to, by id, and they are rechecked on every message.
Tools called through MCP are the app's own tools. Money-moving ones go through the same approval policy as in the browser, so connecting an assistant opens no side door. Setup for each assistant is in ChatGPT and Claude.
What a server publishes
A name and description for each tool, the arguments it takes and the result it returns. The assistant reads these and decides when to call them. Because the schema is explicit, a tool that takes a request id and an answer can be checked by the server before it does anything.
Questions
Can an MCP client spend without approval?
No. In Atta the same policy that gates the browser agent gates every tool call, whichever client sent it.
Related
- Buy from ChatGPT with AttaAdd Atta to ChatGPT as an MCP connector. Search, add to cart and check out as your Atta member, with the same approval policy applied to every purchase.
- Buy from Claude with AttaAdd Atta to Claude as a custom connector or give Claude Code a token. Search, build a cart and check out as your member, under the same approval rules.