SafeCapture MCP

Inspiration

Camera glasses went mainstream faster than the rules around them. Meta sold more than seven million smart glasses in 2025, and IDC forecasts that display-less smart glasses will reach roughly 13.6 million units in 2026, growing to 27.3 million by 2030.

The people in front of those cameras never agreed to be filmed:

72% of Americans worry about being secretly recorded by camera-equipped wearables in public, and only 20% are very confident they could recognize smart glasses.
Only 4% of smart-glasses owners say they always tell nearby people when they are recording.
Two Harvard students built I-XRAY on Ray-Ban Meta glasses, automatically going from a stranger's face to their name, phone number and home address.

The response so far has been bans. Cruise lines such as MSC and Royal Caribbean restrict smart glasses on board, Philadelphia's courts banned recording eyewear from courthouses, UK venues including Soho House and ATG Theatres have banned filming with them, and health systems are prohibiting unapproved smart glasses in their facilities. Privacy is also holding back adoption: 56% of people unlikely to buy smart glasses cite privacy concerns.

Banning the device throws away everything valuable about it: hands-free guidance for a nurse, a field technician or a person with low vision. Recording lights do not solve it either, because most people cannot recognize the glasses in the first place. We set out to make the camera itself respect people, so the glasses can stay on.

What it does

SafeCapture is a programmable privacy layer for camera glasses. It sits between the camera and everything downstream (the screen, the recording, the livestream) and removes what should never be captured, before it is saved or streamed.

Every person we spoke to wanted different rules. A vlogger wants bystanders and licence plates removed; a doctor wants documents, X-rays and children's faces removed; a business owner wants every face and screen blurred except for staff who opt in. So rather than shipping one fixed filter, SafeCapture lets anyone write rules in plain English:

"Blur faces, but let people who show two peace signs unblur, so they can consent to be photographed."

That sentence becomes a live, tested profile on the glasses in about one second.

Private by default, fail-closed. Every face is blurred the moment it appears. If tracking is lost, the region stays blurred, and the entire frame is blurred during fast camera motion.
Consent by gesture. Two peace signs to be seen, two thumbs-down to disappear again. Consent is held only in memory and is wiped when the session ends.
Rules anyone can write. Wearers speak to their glasses ("also blur screens"), business owners type a policy into a dashboard, and AI agents use the same five MCP tools.
Test before you trust. Any profile can be evaluated on synthetic moving-camera footage, so no real person's video is ever required.
Any camera. Runs live on Mentra Live glasses, accepts webcams and network streams, and was prototyped on Meta Quest 3.

How we built it

Claude / any MCP client · web dashboard · voice on the glasses · typing
                            │
                 MCP tools (one interface)
                            │
   profiles/*.yaml  (validated, tested, hot-reloaded in ~1 s)
                            │
 detect → track → consent memory → fail-closed blur
                            │
 in: Mentra Live (USB), webcam, any stream    out: screen, recording

Glasses pipeline (Mentra Live)

Detection: YOLOv8 for people and screens, InsightFace for faces, and MediaPipe for consent gestures.
Rules engine: YAML profiles, validated and hot-reloaded in about one second, exposed through an MCP server with five tools: list_capabilities, save_profile, test_profile, activate_profile and get_profile.
Voice: glasses microphone → local Whisper → Claude with MCP → live profile update.
Dashboard: profile management, a plain-English rule editor, test results and recording controls.

Headset prototype (Meta Quest 3)

Unity 6 with the Meta XR SDK and the Passthrough Camera API.
YOLOv8n running on-device with Unity Sentis, with Meta SAM 3.1 providing open-vocabulary verification through a lightweight Node.js relay that keeps the API key off the headset.
Push-to-talk voice commands transcribed by Meta Muse and parsed on-device into mask and unmask rules ("hide the laptop", "blur this").
Pixelation applied to the raw frame before JPEG encoding, so private content never reaches the stream.

Measuring privacy

We score every profile on synthetic clips using a leak rate:

$$ L = \frac{\text{frames in which a protected face is not fully covered}}{\text{total frames}} $$

Across three moving-camera clips (660 frames), $L = 1.8\%$, concentrated in fast camera motion, with zero people wrongly unblurred. On a synthetic 50-face crowd, leaks dropped from every frame to just two or three frames.

Challenges we ran into

Camera motion. Head-mounted cameras move constantly, and single-frame detection leaks faces. We added hold-and-track logic (regions stay blurred for 0.8 s after a missed detection, padded by 15% and smoothed) and full-frame blur during fast pans.
Consent without surveillance. We needed to remember who opted in without building a face database. Consent lives only in RAM, only for people who opted in, and is erased when the session ends.
Honest limits. When a rule requests something the engine cannot yet do, SafeCapture reports it instead of silently pretending.
Hardware and networking. We worked through native camera crashes and blur flicker on the Quest, campus Wi-Fi that blocked device-to-device traffic, and getting a stable feed from the glasses into the pipeline.

What we learned

Privacy is not a single setting. A hospital, a street and a livestream each need different rules, which is why the rules have to be programmable.
Natural-language rules must be testable. Synthetic footage let us validate privacy behavior without collecting anyone's face.
Voice is the right interface for display-less glasses. Speaking a rule and hearing it confirmed is faster and more natural than any menu.
Fail-closed is a design principle, not a feature. When the system is uncertain, it must protect people first.