serverA modern X11 server for macOS, built in 30 days by a seasoned programmer and Agentic AI
Technical writeups on specific design decisions in macXserver. Bridging an MIT-era X11 server architecture to modern macOS, one constraint at a time.
Each post here picks one specific design decision macXserver had to make and walks through it: what the constraint was, what I tried first, what the right answer turned out to be, and why the code looks the way it does. These are longer and more focused than the 30-Day Sprint entries — the entries say what happened that day; the deep dives explain why one particular thing is built the way it is.
Subscribe via the RSS feed if you want updates as posts land.
X11 is the windowing system that came out of MIT in 1987 and never really left. Here's where it came from, what it solved, why mid-80s PCs were a disaster, and why a 39-year-old protocol still ships on most Linux desktops in 2026. Part of a 30-day experiment building production software with agentic AI.
A live reference for the 119 core X11 protocol opcodes plus the era-correct extensions that real X11R5/R6 clients ship against. Each row says what the opcode does on the wire and what macXserver does with it today. Part of a 30-day experiment building production software with agentic AI.
An honest weighted comparison of macXserver, MIT X11R6, and XQuartz. Parity, wins, losses, and the categories that aren't in macXserver's mission at all.
I thought xcalc would be one of the easy apps. It was easy to get it to work but then I started looking for optimizations that weren't there.
A 1-bit window mask at X logical resolution renders blocky stairsteps on Retina. macXserver converts the mask to a CGPath and lets Core Graphics rasterize it at device scale. xeyes and oclock look round.
A discussion with Claude about generating code from concept vs using known working code.
Motif menus were popping up in the wrong place. The cause was one missing protocol event with a send_event bit set. ICCCM 4.1.5, in detail.
The `reference/` directory in the project tree grew to about 650 megabytes by the end of the project: X11R6, X protocol spec, ICCCM, XQuartz, CDE, Motif, plus a bunch of smaller specs. We built it on Day 2 and should have done it on Day 1. The Motif part was a surprise. I assumed it wasn't available; Claude went looking, found it had been open-sourced over a decade ago, and was already reading it when I noticed.
Why macXserver scales only by integer factors and turns off antialiasing on shape fills. xclock taught us this one on Day 3.
Day 10 was the XError-honesty sweep. Day 11 Claude was about to start bandaiding opcodes one at a time for each failing client. I stopped him and said sweep the whole opcode family from the X11R6 reference at once. Three families fixed in a day. Four 'good-enough' shortcuts caught.
Around Day 9 we flipped the policy from 'silently drop what we can't handle' to 'emit the spec-correct XError every time.' About forty silent-drop bugs surfaced in the next two days.
XQuartz makes every X window into an NSView, nested in AppKit's view hierarchy. macXserver only puts the X top-level into AppKit; descendant X windows are virtual surfaces with their own clipList. The trade-off: we re-implement X clipping in our own code, in exchange for spec-faithful X semantics independent of AppKit.
An xterm names a cell. The server picks a font. Whose metrics win? Walking through the three-day arc that ended at iTerm2 quality on Day 5.
A one-line CTM shift in user space fixed every stroke ghost on every cursor draw. X11 uses pixel-center coordinates; Core Graphics uses grid-line coordinates; the offset between them is exactly half a pixel.
Day 1 wasn't writing the server. It was writing the capture tool. Why we built the framer and replay loop before a single opcode handler, and how the corpus of captured X sessions became the regression test suite for the rest of the project.