Launching today
AI Eyes
Permission-first real-time senses for AI companions
21 followers
Permission-first real-time senses for AI companions
21 followers
AI Eyes explores what it would take for AI companions to share the moment with you—not through hidden monitoring, but through explicit, visible permission. Choose the exact screen or window, control screen, audio, and microphone independently, invite named companions, and pause or end sensing anytime. Built with Codex and GPT‑5.6, v0.1 proves the local-first capture, presence, and session-control layer; live semantic understanding is the next milestone.
splitting mic and system audio controls makes sense on paper but some apps don't even expose that separately, how are you handling those cases.
@ella_cooper We treat microphone and shared-source audio as separate tracks. The mic is requested independently, while system or tab audio is only enabled if the browser and selected source actually provide an audio track. If they don’t, AI Eyes disables that control and shows an explicit “audio unavailable” state instead of pretending it’s captured. Once tracks exist, either can be muted or stopped independently. Cross-browser support is still one of the rough edges we need to test properly.
real question is whether re-granting access every session feels annoying or feels safe, which way's it leaning so far for people testing it.
@stella_reed1 The current prototype requires the user to select the source again for each new session. That’s intentionally the safer default, but it’s too early to claim whether users find it reassuring or annoying. A future version could remember preferences without retaining silent capture access. I’d rather earn convenience gradually than make persistent access the default.
shadow's kinda doing his already no? what's actually different about the permission side here vs theirs.
@sadie_charlotte Fair question—there’s definite overlap with Shadow. Shadow is already a polished Mac product focused on meetings, screen context, and actions. AI Eyes v0.1 is narrower: it explores explicit session control for named AI companions—exact source selection, independent screen/system-audio/mic controls, visible sensing state, and temporary-context clearing when the session ends. We haven’t surpassed Shadow’s semantic layer; we’re testing a different starting point.
getting session control and local-first right before the flashy semantic stuff is smart tbh, most people rush the other way first.
@violet_clark1 Thank you—that ordering is intentional. Before an AI understands anything, I want the human to see exactly what is being sensed, control every source independently, and end the session cleanly. The semantic layer comes next.
is the actually usable rn or just a demo?
@ayazakram Honest answer: it’s currently a functional permission and session-control prototype, not a complete AI companion. Screen capture, source controls, pause/end behaviour, and context clearing work; semantic understanding and agent responses are still simulated. The next milestone is connecting the real multimodal layer.
yeah permission-first kills some of the "ai just gets it" magic but ngl it's the only version of this i'd actually let near my screen.
@daniel_carter18 That trade-off is exactly what I’m exploring. Some magic disappears when every sensing state is visible—but if the alternative is hidden or ambiguous access, I’d rather make the controls clear and then carefully earn back the convenience.