Try this in whatever AI assistant you already have open. Ask it what you should wear to dinner on Thursday. It will give you an answer — a confident, well-written, completely fictional answer about a navy blazer and a crisp white shirt, neither of which it has any reason to believe you own.
That is not the model being dumb. It is the model being asked a question it has no way to answer. It knows an enormous amount about clothes in general and precisely nothing about your clothes in particular. The gap between those two things is the entire reason "ask ChatGPT what to wear" has never actually worked.
The Missing Half of the Question
Getting dressed is a question with two halves. The first half is general knowledge: what reads as smart-casual, which colors sit well together, how formal a garden wedding really is. Language models are genuinely excellent at this half. They have read every style guide ever published.
The second half is inventory. What do you own, what fits, what is clean right now, what did you already wear on Monday in front of the same people. No amount of general knowledge produces that. It is a fact about your bedroom.
So an assistant answering from the first half alone has to invent the second half, and it does — fluently, plausibly, and uselessly. You cannot wear a recommendation. You can only wear a garment.
What Changed: A Protocol for Handing Models Real Context
The Model Context Protocol, usually shortened to MCP, is an open standard for exactly this problem. It gives an application a way to expose a small, deliberate set of capabilities — read this, search that, do this specific thing — that an AI assistant can discover and call during a conversation. The assistant stops guessing and starts asking.
It is worth being clear about what it is not. It is not the assistant being given the run of your account. An MCP server publishes a fixed list of tools, each doing one named thing, and nothing outside that list is reachable no matter how the conversation goes. The interesting design work is in deciding what belongs on the list — and, more importantly, what never does.
What This Looks Like for a Wardrobe
Clad9 runs an MCP server, so Claude, ChatGPT, Gemini or anything you have built yourself can reach the closet you have already cataloged. The conversation stops being hypothetical:
- "What navy do I own?" — a search across your actual garments, in plain language, not a list of navy garments that exist in the world.
- "What's in the wash?" — your laundry basket, and how much of your wardrobe is currently in it.
- "Something for a client dinner." — three complete looks built from clothes you own, with the picture, so you can see the outfit rather than read a description of it.
- "Same look, different shoes." — one piece re-picked against the rest of the outfit, and the outfit redrawn.
- "Dress me for Thursday's dinner." — the actual event on your actual calendar, with its dress code and its weather.
- "I wore that." — logged as worn, with the washable pieces sent to the laundry basket.
The useful shift is subtle but total. The assistant is no longer performing the role of a stylist. It is talking to one, on your behalf, in the window you already had open.
One Product, Two Doors
There is a failure mode worth naming, because plenty of products fall into it: the API version of the product slowly becomes a different, worse product. Different limits, different rules, a recommendation engine that behaves one way on the website and another way over the wire.
The rule we set for ourselves is that MCP is a second door onto the product, never a second product. An outfit asked for in a chat window goes through exactly the same code as one asked for on the website. It spends the same daily allowance. It skips the same laundry basket. It obeys the same rule about not repeating what you wore on Tuesday. Nothing re-derives a limit or a refusal on the other side of the door — because two implementations of the same rule eventually disagree, and the day they disagree is the day the product starts lying to somebody.
The Part That Matters Most: What It Will Never Do
Connecting an AI assistant to something personal deserves more scrutiny than connecting it to a to-do list, and a wardrobe app holds photographs. So the boundaries here are not permissions you can grant and choose not to. They are not offered at all.
- No photographs of you, ever. Selfies, body photos and fit-check pictures are never sent to a connected assistant, at any permission level, in any form. It can be told which garments a fit check matched. It cannot be given the picture. Photos of your clothes, and the outfit images Clad9 draws, can be shared — that is what makes the answer useful. A photograph of you is a different thing.
- Nothing destructive. No deleting garments, no plan or account changes, no payment details, no disconnecting your calendar.
- No adding to your closet. Cataloging a garment is a photo you took deliberately. An assistant looping over that is a bill and a polluted wardrobe, so capture stays something a person does.
- A log you can read. Every request an assistant makes is recorded and shown back to you in plain language — "Asked for an outfit for 'client dinner'", "Sent 3 items to the laundry basket". A log you cannot read is not consent.
That last one deserves a sentence of explanation. An assistant's context can hold a web page, an email, or a calendar invite written by somebody who is not you. Designing for that means assuming the instructions reaching the tools may not have come from the account holder — which is an argument for a short tool list, no destructive actions on it, and a visible record of everything that happened.
The Honest Caveat
One thing is genuinely outside our control and worth saying plainly: once your question and Clad9's answer reach the assistant, they are handled by whoever makes that assistant, under their privacy policy and their retention and training terms. We cannot delete it for you. Read their policy before you connect them — it is a real decision, not a formality.
Worth Doing?
If you already live in an AI assistant all day, this removes the last reason your wardrobe was not part of that. You stop switching apps to ask a question you were already mid-conversation about. If you don't, you lose nothing — the Clad9 app answers the same questions the same way, because it is the same engine underneath.
Either way, the prerequisite is the same and it is not the protocol. An assistant can only answer from clothes Clad9 knows about. Photograph your closet first, let Clad9 identify what is in it, and then point whatever you are already working in at it.
