What it doesCopy Anchor Link

In Faces and ports, you can keep several photos of a device, trace a face with any of them, and ask the modeler to find ports on the photo. It marks what it finds as candidates: port openings by kind (USB-A, USB-C, RJ45, HDMI, SFP, QSFP, a DC jack…), vents and grilles, LEDs and buttons, each with how sure it is. You accept, correct or reject them, and so can your AI agent.

Adding photosCopy Anchor Link

Under the face you're working on, Photos of this device keeps as many photos as you like (up to 40 a device):

• drop photos onto it, or choose Add photos… to pick several files;
• on a phone, Take a photo opens the camera;
• each photo belongs to a face, or stays unsorted. New photos go to the face you're on; change it with the menu next to the photo.

Trace the front face with it puts a photo under the face. Click its four corners, as for any traced photo (see Tracing photos and drawings): the photo is straightened and scaled to the face, so every pixel maps to millimetres. The corners are remembered with the photo, so switching back to it later needs no new clicks.

Photos are downscaled to 2,000 pixels on their long side and kept in this browser's storage. They follow the device when its draft is first saved.

A photo you traced with Photo…, one traced before the device's photos existed, and one that reached this browser from your other browser by live sync are listed here too, with the face they're traced on and their corners. They are copied into the list once, in this browser; nothing is uploaded.

Finding portsCopy Anchor Link

Once the face's corners are placed, choose Find ports on this photo. It takes a fraction of a second, in the background, while the page stays usable. Candidates appear as dashed teal boxes on the face and on the photo, labelled with their kind and confidence, e.g. d3 usb-c 86%.

How it works: the face is straightened at up to 10 pixels per millimetre; openings, holes and lit or coloured dots are found against the face's own colour; a group of alike holes or slots becomes one vent; and each other shape is compared with the standard openings in the modeler's port library, by size and outline. There is no downloaded model: it is a few kilobytes of code.

Detection needs a signed-in account and a saved device (your draft, or the published device you opened), since what it finds is kept with the device. Until it can run, the button says what to do first, and so does the line below it: for example Trace the corners of the front face first, or trace the face with one of its photos. Running it again on the same face replaces the candidates still to review, and keeps the ones you've already accepted or rejected.

Reviewing candidatesCopy Anchor Link

Click a candidate to review it below the face:

• What it is: pick any kind, or one of the detector's next best guesses in one click;
• u, v, width, height: its box in millimetres. You can also drag the box on the face to move it, or drag its corner to resize it;
• a note for your agent ("this is the PD input", "check this one");
• Accept makes the port, zone or LED with the right kind and size: a port stood on end is turned, and one whose opening differs from the standard size keeps its own. It's an ordinary edit: it undoes, saves and shows in your other browsers;
• Reject drops it (a screw, a label, a shadow); Back to pending undoes a review.

Accept all above 80 % (or any other threshold) accepts every candidate on the face that sure of itself, as a single undo step. Openings of no known kind are left for you to name.

A row of identical ports is found one by one. Accept the first, then set its count and pitch in the port's form, and reject the rest, or ask your agent to.

With your AI agentCopy Anchor Link

Your agent reads the candidates as numbers, never the photo, and uses them together with what you tell it in the chat and your annotations:

• get_draft tells it how many candidates are waiting, and how many are unsure;
• list_detections gives each one's kind, confidence, box in millimetres, the next best kinds and your notes;
• its face drawings (render_faces) show the waiting candidates as dashed boxes, apart from your annotations;
• it accepts (accept_detection), rejects (reject_detection) or relabels (update_detection) them, and you see each change at once in the modeler.

For an uncertain candidate it should ask you first, with a measure this annotation where the candidate is, or a reply to your note. See the MCP tool reference.

LimitsCopy Anchor Link

• Detection works on flat faces traced from a photo; not on the round side of a cylinder.
• It finds openings and dots that stand out from the face. Ports in a dark recess on a dark body, glossy reflections and printed logos can be missed or mistaken: review what it finds.
• Kinds that look alike on a photo, such as HDMI and DisplayPort, may come with a lower confidence and the other kind as the next best guess.
• Your photo library stays in the browser where you added it. The photo a face is traced with reaches your other open browsers through live sync as before, and joins their photos; the other photos don't.