Open-source r2d2-tv turns webOS TVs into privacy-first smart displays with local voice assistant

An independent developer has published r2d2-tv, an open-source project that rebuilds LG webOS televisions as privacy-oriented smart displays. Shared on Hacker News through its GitHub repository, the software supplies three things: a local home screen, a browser rendered on the user's own hardware, and a push-to-talk voice assistant whose speech models run on the local network rather than in a vendor's cloud.

How the pieces fit together

According to the repository, the system splits responsibilities between the TV and a separate computer. The host — an aarch64 Linux machine — renders a sandboxed Chromium session inside a private display and audio environment and streams it to the TV over the LAN. The same machine runs the speech recognition and intent models, along with a closed registry of commands the assistant is permitted to execute.

The television keeps its role as a screen, app launcher, and input device. A Home application and a browser viewer run on the set, with the viewer relaying remote-control input back to the host. For voice, a root-level capture component takes ownership of the remote's voice key and forwards a bounded audio stream to the host over mutual TLS.

Content blocking is handled through pinned filter rules loaded into each disposable browser profile, an approach similar in spirit to desktop ad blockers.

Voice inference stays local

The project's privacy pitch is specific: model inference never leaves the host machine. The repository is equally clear about the limits of that claim — requested websites, searches, streaming applications, and an optional weather feature still contact their own providers. A SECURITY.md file documents the boundaries.

Demanding prerequisites

Running the stack is not trivial. The documented host requirements include a non-root service account, the Chromium snap with its sandbox enabled, Xvfb for a private X display, PipeWire, WirePlumber and GStreamer plugins for the WebRTC path, FFmpeg with hardware HEVC encoding for the MSE and AirPlay paths, and a CUDA-capable GPU with Python 3.12 to serve the pinned local models — roughly 10 GB of weights.

On the TV side, the software assumes system-level root access to a webOS set and an authorized way to install packages. Native voice capture additionally requires a toolchain and headers matching the exact firmware build. Builds are produced locally and do not contact the device; the browser shortcut is gated behind a private capability key, and brand logos are not shipped, leaving text tiles on the home screen unless the user supplies artwork they have the rights to use.

Explicit risk warnings

The README is unusually blunt about the dangers. The TV components run as root and replace or hook vendor system services, which the author describes as unsupported, potentially warranty-voiding, and capable of leaving the TV unusable until a reboot or a factory reset. Firmware hashes, code offsets, symbol names, and patch anchors all target a single reference build; users must re-derive them for their own firmware and are told never to force a mismatched pin. A mistake can disable the remote, the Home button, or voice input until the set is restored.

The status section narrows expectations further: the end-to-end configuration — Home, the browser, and remote voice commands — has been validated on one TV and firmware combination. Every other mix of television, firmware, remote, and host is untested, and the repository offers source code and offline tests rather than a universal installer.

The project is MIT-licensed, with the browser extension, filter tooling and data, and two runtime adapters under GPL-3.0-or-later. It is fully independent, with no affiliation to LG Electronics, Lucasfilm, or The Walt Disney Company.

Why it matters

Most smart TVs funnel their home screens, search, and voice features through manufacturer cloud services. r2d2-tv demonstrates a working alternative that keeps the panel and app ecosystem but relocates the brain — home screen, browsing, and voice — to hardware the owner controls, with local model inference and LAN streaming. It is a hobbyist blueprint rather than a consumer product, and the price of admission is steep: rooted firmware, build-specific binary patches, a CUDA GPU, and 10 GB of model weights. Even so, it shows what a privacy-first smart display can look like once the assumption that the intelligence must live in someone else's datacenter is removed.