PraeterVNC

A performance focused VNC (RFB 3.3/3.7/3.8) server for Windows 10/11, written in Rust.

How To Run?

• Portable: run PraeterVNC-<ver>-portable.exe. Settings go in praetervnc.ini next to it (%LOCALAPPDATA%\PraeterVNC if that folder is read-only).
• Service: install the MSI, or use tray menu -> Install as service… (or praetervnc --install). Runs as a SYSTEM service & starts a helper. Settings from the portable copy carry over.
• Console: praetervnc [options] runs the server without a tray.

Option	Default
--port N / --bind ADDR	5900 / 127.0.0.1 (0.0.0.0 with a password)
--password PW	none	VNC authentication (DES; tunnel it on untrusted networks)
--monitor N / --list	all monitors	capture one output / list outputs
--view-only	off	ignore input and client clipboard
--inflight N	32	max updates in flight
--alr-ms N	150	idle time before lossless refresh of lossy areas
--budget-ms N	25	encode budget per update
--no-scroll --no-motion --no-pipeline		disable scroll detection / motion JPEG / legacy pipelining
--verbose	off	per-client stats every 5 s

Requires an AVX2 CPU and Windows 10/11 x64!

Security

• Without a password Praeter listens on 127.0.0.1 only.
• Failed logins are counted per IP. Blocked IPs show in the tray.
• VNC auth is DES challenge/response and the session is unencrypted.

Recovery

• A watchdog tracks capture, encode, and session loop. Any problem is logged Praeter restarts itself or is restarted by the service (helper).
• The service pings the helper via the control pipe, and will restart it if it stops answering.
• Release stuck keys sends key-up for any held modifier and mouse button.
• TCP keepalive (10 s) and a 30 s write timeout autodrops dead viewers.
• Logs: logs\praetervnc.log. Rotated at 2 MB.

Build

cargo build --release          # target\release\praetervnc.exe
python tools/package.py        # dist\: MSI (WiX 5, repo-local dotnet tool), portable exe, licenses, SHA256SUMS

Design

Capture

• DXGI Desktop Duplication w/ one thread per output (MMCSS, GPU priority).
• Dirty and move rects are 64 px tiles diffed in parallel against immutable Arc snapshots.
• Per-client tile model: dirty = Arc pointer inequality. (All clients share one capture)

Encoding

• Tight (fill, mono, palette, zlib, JPEG), ZRLE, TRLE, JPEG-21, Hextile, Raw, CopyRect.
• Rich + alpha cursors, PointerPos, DesktopSize/ExtendedDesktopSize, LastRect, Fence, ContinuousUpdates, QEMU key events, ExtendedClipboard.
• Parallel zlib in one valid stream (dictionary primed chunks, sync flushed)
• Scroll detection -> CopyRect.
• Motion tiles encode w/ cheaper JPEG (4:2:0)

Scheduling and flow control

• Per update: urgent work (small changes, near pointer, scroll) -> oldest work within the encode budget -> lossless refresh.
• Motion JPEG quality is determined via link speed.
• Updates are never queued.
• Large changes and lossless refresh are split into chunks.
• Below ~100 Mbit: more zlib effort - ZRLE instead of TRLE.

Results

See BENCHMARKS.md.

Tooling

• crates/testapp: test window / benchmark of scenarios (ticker, typing, scroll, drag, video) with a frame-ID barcode.
• crates/bench: headless RFB client. Reports frames, latency, bandwidth, paint time and server CPU.
• crates/netem: TCP proxy with delay and a bandwidth cap for artificial network latency.
• PRAETER_PROBE=1: per-stage latency.

License

MIT, see LICENSE. Third-party licenses: THIRD-PARTY-NOTICES.txt. praetervnc --licenses prints both.