bring-your-own-agent
Bring your own agent
Give an app somebody already has an assistant that can do its real work — look things up in it, add things to it, answer questions about what is in it.
The result is one new folder inside their project, called agent-access/. It holds a small program that speaks MCP — the standard way an assistant like Claude connects to something and is handed a short list of things it is allowed to do. That program reaches the app over the web, through the doorway the app already offers other programs — its API — so the app's own checks and rules apply to everything the assistant does, exactly as they would to a person clicking around in it. Not one file the owner wrote is changed, and deleting the folder puts the app back exactly as it was. Say that to them in those words; it is the difference between this and a change to their codebase.
The trade that buys it is real and has to be said out loud: the assistant acts as one fixed login. Anything that login can see or change, the assistant can see or change too. There is no permission screen and no way to give one person less access than another. Some apps can do far better than that, which is why Step 0 comes before anything is built — as soon as you know what the app is made of, and before a single word is said about what the assistant might be allowed to do.