Create a token that lets Claude, Cursor or another AI tool work on this app or on all your apps, choose what it may do, watch its activity and revoke it.
MCP is a common way for AI tools such as Claude Desktop, Cursor or Codex to work inside other services. With it you can write to Claude "add a waiting list to my class page and publish" and it does the work in your Fine Structure app. To allow that, you give the tool a token: a long private code that works like a key, with limits you choose. The MCP Server tab of an app is where you make these keys, see what they did, and take them back.
Screen recording: Settings, MCP Server: Agent Access.
1. Open the tab: In the Studio, open Settings and choose MCP Server. The screen is titled Agent Access.
2. Choose the reach: Current app only: locked to this app, and it cannot create new apps. All my apps: can work with the apps you own and create apps if allowed.
3. Choose what it may do: Under Token permissions, press Read only (look, change nothing), Build (change and build) or Full access. You can also tick single permissions (scopes, the list of things the key may do).
4. Create the token and copy: Press the create button (Create app token, or Create account token for all apps). Then One thing to copy. Paste it into the AI/MCP client. Paste it in the connection settings of Claude or Cursor.
Treat a token like a password: Anyone who has it can do what its permissions allow. Paste it only into your own AI tool, never in a chat or an email. If it leaked, press Revoke and make a new one.
Further down, Advanced app-local MCP files can write MCP helper files inside the app itself, so the app you built becomes a tool for AI agents. Create MCP Server from existing system asks the builder to create the mcp/ folder; it reads MCP files exist when they are there. This is for developers; most owners never need it.
Q: Do I need this to use Claude with Fine Structure?
A: The account-wide connection is easier for most people. See Connect AI tools. This tab is for a key limited to one app, or for tools that need a token.
Q: Which permission should I pick?
A: Start with Read only to try it, then Build when you want it to change things. Full access only for tools you fully trust.
Q: Does the tool see my secrets?
A: No. Secret values are never sent to AI tools; they see only the names.
Q: Can I see who changed what?
A: Recent MCP activity lists it, and every change keeps a saved version. See Saved versions.
Related: Connect AI tools: Claude, ChatGPT, Cursor and more, Studio settings, section by section, Secrets: keys for outside services, Saved versions and restore