NUDE Private Fitting Room

Your agent prepares the fitting room. You make the final choice.

Comment

About the project

Most virtual try-on tools are designed to help people check out faster.

We wanted to build something different: a private lingerie fitting room that a person's own agent can prepare for them, without taking the final decision away from the person.

Lingerie shopping is unusually personal. Fit, comfort, color, recovery needs, body confidence, and privacy can matter as much as product specs. An agent can do a lot of useful work — research products, compare options, remember constraints, and prepare a shortlist — but there is a point where automation should stop.

That became our core principle:

Your agent prepares the fitting room. You make the final choice.

What we built

NUDE Private Lingerie Fitting Room combines a real virtual try-on experience with a WebMCP layer that lets a user's own agent operate the shopping workflow.

Through WebMCP, the agent can:

  1. Research available pieces.
  2. Inspect structured product details.
  3. Read the current fitting-room state.
  4. Request a virtual try-on after the user gives consent.
  5. Verify that the correct product asset was delivered.
  6. Prepare a fitting room with recommended pieces and reasons for the user to review.

The final choice is deliberately left to the person.

In our demo, the user is six weeks into shoulder rehabilitation and wants options that are easier to manage while recovering. Her agent researches the collection, proposes pieces, asks for consent before initiating virtual try-on, and prepares three completed looks.

The agent initially chooses the wrong color. The user simply says:

“i need the beige one”

The agent corrects itself and continues.

That moment matters to us. The goal is not an agent that pretends to be infallible. It is an agent that can do useful shopping work, accept human correction, and hand the decision back gracefully.

The agent corrects. The person decides.

How WebMCP fits

The fitting-room product existed before this challenge, including the NUDE shopping experience and Perfect Corp virtual try-on.

During the submission period, we added the WebMCP layer that makes the workflow operable by a user's own agent.

We expose five real tools:

  • nude.list_pieces
  • nude.get_piece
  • nude.get_fitting_room_state
  • nude.try_on
  • nude.prepare_fitting_room

This turns a normal website interaction into a structured agent workflow:

Research → Inspect → Try on → Verify → Prepare → Human handoff

For us, the interesting part of WebMCP is not reducing the number of clicks. It is allowing an agent to do meaningful preparatory work while the website remains the place where the person sees the result and makes the final judgment.

Privacy and the data boundary

Privacy was especially important because this is a lingerie fitting experience.

The user's photo is sent to Perfect Corp when virtual try-on is requested. However, no photo of the person and no try-on result is returned through the WebMCP tools.

The agent works with structured information such as product data, the user's brief, fitting-room state, completion status, and verified product delivery.

This also means our recommendation reasoning is intentionally grounded in the user's brief and structured product information, not in pretending that the agent visually inspected the person's try-on image.

The rendered result is there for the person.

Challenges we faced

The hardest part was deciding what the agent should not do.

It would have been easy to make the demo look more magical by implying that the agent could see everything, judge the person's body, or make the final selection automatically. We chose not to.

Instead, we had to design clear boundaries between:

  • what the agent knows,
  • what the try-on service renders,
  • what WebMCP returns,
  • and what only the person should judge.

We also had to make the workflow robust enough for correction. If the agent prepares the wrong color or misunderstands a preference, the product cannot collapse. The user must be able to correct it naturally and continue.

Finally, we learned that agent UX is not simply normal UI with an API added. The agent needs structured state, clear tool semantics, verifiable completion, and an explicit handoff back to the human.

What we learned

The biggest lesson was that good agentic commerce is not maximum automation.

An agent can remove a lot of shopping work without removing human agency.

WebMCP gave us a way to separate those two things cleanly: the agent can research, operate tools, verify state, and prepare the experience, while the person remains responsible for the subjective decision that actually matters.

For a category as intimate as lingerie, that balance feels more useful than an autonomous checkout button.

We think the future of shopping agents is not:

“The AI chose for me.”

It is:

“Everything is ready when I'm ready to choose.”

What people and agents can now do together

We gave the same brief — ‘’ six weeks into shoulder rehab, can't reach behind her back, needs something invisible under white office shirts ‘’ — to an agent with no site tools, and to the same brief through WebMCP.

Without tools, the agent produced a warm, well-organised consultation. It recommended a specific piece, described its cup as "seamless" — that is a different product in the catalogue; this one is two-panel seamed and push-up — asserted the nude colourway would be "completely invisible" under a white shirt, then asked her to simulate the fitting in her mind and to run the white-shirt test herself, on a bra she does not own.

Through WebMCP, the same brief produces three real Perfect Corp try-ons on her own photo, a per-piece reason the agent had to supply before the fitting room would accept it, and this caveat: "VTO cannot prove white-shirt invisibility or all-day comfort; confirm permitted movements with the rehab clinician."

The tool-less answer reads better. It is warmer, more confident, better structured. That is the problem. Fluency is not verification, and this decision lands on a body — someone in week six of rehab, deciding what she can physically put on. WebMCP is what lets the page hand the agent the things it can actually check, and refuse the things it cannot.

Built With

  • corp
  • css
  • framer
  • indexeddb
  • motion
  • next.js
  • perfect
  • react
  • tailwind
  • typescript
  • vercel
  • virtual
  • webmcp
  • youcam
  • zustand
Share this project:

Updates

Submission history