← Otto Developer GuideAirPlay video screenshare — design exploration

Status: exploration; nothing implemented in Otto. Phase 0 below — sending a virtual output to an Apple TV via doubletake, with no Otto code at all — has been validated against real hardware. Everything past that is still design space.

This doc maps that space in both directions and recommends a path. It follows the otto-rdp pattern (rdp-virtual-output.md): a standalone component that consumes a virtual output’s PipeWire node, with the compositor knowing nothing about the remote protocol.

Two directions, two features

	Sender (Otto → Apple device)	Receiver (Apple device → Otto)
What it gives	Apple TV / AirPlay TV becomes an extra (virtual) Otto display	iPhone/iPad/Mac screen appears as a window in Otto
Input backchannel	none — AirPlay mirroring is display-only	none
Protocol difficulty	hard (FairPlay SAP on the sender side)	solved (UxPlay, RPiPlay, shairplay)
Prior art	doubletake (Go, active), AirplayMirroringGo (dormant)	UxPlay (GPLv3), shairplay Rust crate (LGPL, mirroring experimental)

Unlike RDP there is no input path in either direction — AirPlay mirroring has no HID backchannel a third party can use. So the RDP bridge remains the “remote-controllable screen” story; AirPlay is view-only: interactive = false virtual outputs are the right source, and none of the virtual-pointer/keyboard compositor fixes are needed.

Sender: Otto virtual output → Apple TV

Protocol reality

AirPlay 2 mirroring is mDNS discovery (_airplay._tcp) + RTSP-like control on port 7000 + an H.264 elementary stream (128-byte-header packets, encrypted) on a separate TCP port, with PTP/NTP timing. The gatekeeping steps:

pair-setup / pair-verify — HomeKit-style SRP-6a with a PIN on first contact (mandatory since tvOS 10.2), Ed25519/X25519 session keys, ChaCha20-Poly1305 channel encryption. Fully reverse-engineered; pyatv and doubletake implement it; all the crypto exists as Rust crates.
fp-setup (FairPlay SAP) — the sender must produce a valid FairPlay handshake that a genuine Apple TV accepts. This is the blocker for a clean-room sender: receivers get away with a small reverse-engineered decryptor (playfair), but the sender side has only been done by executing snapshots of Apple’s own ARM64 FairPlay code (doubletake calls this “snapshot-backed Go ARM64 execution”, the Slave-in-the-Magic-Mirror approach). Legally gray, unpleasant to port, breakable by tvOS updates.

The pragmatic observation: doubletake already works with Otto

doubletake (LGPL-3.0, v0.4.0 July 2026) is a Linux AirPlay 2 mirroring sender that captures via xdg-desktop-portal / PipeWire and encodes with VA-API/NVENC. Otto already ships xdg-desktop-portal-otto and exposes every output — including virtual outputs — as portal monitors backed by PipeWire nodes. So in principle:

[[virtual_outputs]]            # otto_config.toml — view-only is fine
name = "airplay-1"
resolution = { width = 1920, height = 1080 }
refresh_hz = 30.0
position = { x = 1440, y = 0 }

$ doubletake   # pick "airplay-1" in the portal dialog, pick the Apple TV

gives “Apple TV as a second Otto monitor” with zero Otto code. Caveats: largely LLM-written by its author’s own admission, security posture unknown, tested against Apple TV 4K / M-series Macs / some TVs (Xiaomi broken).

Native sender (`otto-airplay-tx`) — if we ever want one

Reuses two of otto-rdp’s three modules nearly verbatim:

virtual output PipeWire node ──▶ pipewire_capture.rs (BGRx frames)
        ──▶ H.264 encode (VA-API; new)
        ──▶ AirPlay2 session: mDNS + pair-verify + fp-setup + RTSP  (new)
        ──▶ encrypted 128-byte-header video stream + PTP timing     (new)

Capture + org.otto.ScreenCast glue: exists (pipewire_capture.rs, screencast.rs).
Encoding: new dependency surface. Otto has no gstreamer/ffmpeg today; the lean option is direct libva bindings (or spawning ffmpeg), not a gstreamer stack. 1080p30 BGRx → H.264 at ~10–20 Mbit is the target; AirPlay receivers expect baseline/main H.264, SPS/PPS in-band.
Protocol: pairing/crypto is a week of work with existing crates (srp, x25519-dalek, ed25519-dalek, chacha20poly1305, mDNS, bplist). FairPlay SAP is not — the realistic route is linking or spawning doubletake’s fairplay module rather than reimplementing it. That keeps the gray-area code out of Otto’s MIT tree (LGPL helper process).
Latency: AirPlay mirroring lands around 150–500 ms end-to-end (encode + receiver buffering). Fine for video/slides/reference material on a TV, not for typing at — set expectations accordingly in docs/UI.

Non-mirroring fallback: AirPlay Video (HLS push)

play_url-style AirPlay (what pyatv/open-airplay do) needs no FairPlay for non-DRM content: tell the Apple TV to fetch an HLS URL we serve. We could encode a virtual output (or a single window) into LL-HLS and cast it. Latency is seconds — useless as a display, plausible as “cast this video/window to the TV”. Cheap to build, but a different feature than screenshare; parked.

Receiver: iPhone/iPad/Mac screen as an Otto window (`otto-airplay-rx`)

The reverse direction is well-trodden: advertise _airplay._tcp + _raop._tcp, accept the mirroring session (receiver-side FairPlay is the solved playfair problem), decode H.264, present. Two integration shapes:

Wayland-client component (recommended): each incoming session creates an xdg toplevel and renders decoded frames (VA-API decode → dmabuf → wp_linux_dmabuf). No compositor changes at all; it composes with everything (expose, workspaces, screenshare re-share).
Deeper compositor integration (dedicated layer/scene node) buys nothing the window doesn’t already give.

Building blocks: shairplay crate (LGPL-3.0, pure Rust, AirPlay 1+2, mirroring behind an experimental video feature — evaluate first) or porting UxPlay’s session logic (GPLv3 — cannot be vendored into MIT Otto; process-boundary only). A macOS sender can also use an AirPlay receiver as an extended display, which would make an Otto machine a wireless external monitor for a MacBook — worth verifying against UxPlay first to confirm third-party receivers get that mode.

Recommendation

Phase 0 — validate with doubletake (no code). Done. Virtual output + portal + doubletake against a real Apple TV works: Otto’s capture path carries an AirPlay session end to end with no compositor changes. So the sender direction does not obviously need to be owned natively — the “integration” may amount to documentation plus a dock affordance that launches it. A receiver build and the audio path are still open.
Phase 1 — otto-airplay-rx if receiving is wanted: highest feasibility, pure-Rust path available, and it’s a visible differentiator (iPhone screenshare lands as a normal Otto window; possibly Mac wireless display). Component-only, MIT-safe via LGPL crate.
Phase 2 — native sender only if Phase 0 shows demand and we accept the FairPlay strategy (external LGPL helper for fp-setup). Skeleton is 80% otto-rdp minus input, plus an encoder.