// the find
elder-plinius/V3SP3R
AI Flipper control
Vesper is an Android app that puts an LLM (via OpenRouter) in front of a Flipper Zero over BLE, so you drive SubGHz, IR, NFC, BadUSB and GPIO through chat, voice, or a photo instead of the stock device menus. It's aimed at people who already own a Flipper and want a faster interface — security researchers, red teamers, CTF players — not beginners trying to learn RF or hardware hacking from zero.
The risk model is tiered rather than binary: reads execute automatically, writes show a diff, destructive ops (subghz_transmit, badusb_execute) need a double-tap confirm, and every call is written to an audit log — a sane shape for handing an LLM a radio and a HID injector. It's model-agnostic through OpenRouter instead of hardcoding a vendor, so you can swap in whatever's cheapest or best at tool-calling that week. The subsystem coverage (BLE protocol, file system, SubGHz/Pwnagotchi protocol handling) looks like real integration work, not a UI mockup — there are unit tests for FlipperProtocol command mapping, CommandExecutor, and RiskAssessor specifically, which is where you'd want them.
The repo has build output checked into git — app/build/outputs/apk/debug/app-debug.apk, the whole intermediates tree, .DS_Store files — which means no working .gitignore for the app module and a needlessly bloated clone. It's Android-only; iOS is listed under 'needs love' with nothing built. Every single action, including a battery check, requires a round trip to a paid cloud LLM, so the tool is fully non-functional offline and adds real latency/cost to what's pitched as an 'instant' interface. The actual safety boundary is a confirmation dialog gating whatever the model decides to call — nothing in the README addresses what happens if a downloaded 'resource' or an analyzed photo contains content designed to steer the model into proposing a high-risk action, and consent-fatigue on repeated double-taps is a real failure mode for exactly that kind of design.