Two Per Cent of AutoCAD
I'd just asked Claude a daft question: what would it cost me, in Amazon Bedrock tokens, to build an open-source, clean-room copy of the Adobe suite?
The answer was a back-of-envelope guess of somewhere between a quarter of a million and four million dollars, with the caveat that the money wasn't really the problem. Without humans steering, it said, a project that size doesn't converge. It drifts. The bit that kills you is the long tail: the file-format edge cases, the features nobody demos, the decades of work hiding behind every menu.
So I was surprised when a friend sent me a link to ArtCraft.
Someone's already done it
ArtCraft lists thirteen "Crafting Apps": PhotoCraft, VectorCraft, FilmCraft, PdfCraft, EffectCraft, CADCraft and seven more. Photoshop, Illustrator, Premiere, Acrobat, After Effects, AutoCAD, and on through Word, Excel and Pro Tools. The site lists them as open source, written in Rust, and notarised for macOS. CADCraft's README describes itself as "an open-source, clean-room reimplementation of Autodesk AutoCAD, rebuilt in pure Rust."
Better still, every app is "agent-ready": "Drive every app from a CLI, a JSON control channel or an MCP (Model Context Protocol) server, built for automation and AI agents."
So much for Claude's upper estimate of four million dollars. I installed PhotoCraft, EffectCraft, PdfCraft and CADCraft, and decided the right response to my friend was to have Claude drive one of them and send back something impressive. A 3D-printable humanoid figure, say. An .stl (a common 3D-printing format) straight out of their own AutoCAD.
Wiring it up
This part was pleasingly boring. The MCP servers aren't in the app bundles; each app has a separate *-cli download on its GitHub releases page, and photocraft-cli mcp (or cadcraft-cli mcp, and so on) speaks MCP over stdio. All four turned up with their tool lists on the first try:
photocraft 25 tools
effectcraft 27 tools
pdfcraft 17 tools
cadcraft 11 tools
They went in behind meta-mcp, my proxy that keeps nearly twenty MCP servers out of Claude's context until they're needed. Four TOML entries, a reconnect, done.
Then I asked for the humanoid.
No STL, no solids
Claude came back with a short answer: no.
CADCraft 0.6.2 exports DXF, DWG (AutoCAD's native format), SVG, PNG and PDF. There's no STL. There are no solids to export, either. Its only commands with "3D" in the name are 3dpoly, a 3D polyline, and 3dorbit. union exists, but it unions flat 2D regions.
"I thought CADCraft was supposed to be like AutoCAD?" I asked, slightly wounded on my friend's behalf.
It is supposed to be. So Claude went and read the roadmap.
The roadmap is candid. The homepage is optimistic.
Credit where it's due: CADCraft's ROADMAP.md is one of the most candid project documents I've read. It opens with this:
Stage: pre-alpha
It doesn't stop there. As of its last review on 10 October 2026:
| Number | Value |
|---|---|
| Feature breadth (AutoCAD for Mac menu items with a live command) | 50% (245 / 491) |
| Ready for real work | ≈ 38% |
| Mainstream practitioner (2D drafter) | ≈ 28% |
| Essentials user | ≈ 46% |
| Remaining to full parity | ≈ 1,300–2,100 agent-hours |
It moves fast, too: the parity checklist regenerated two days later already said 57% breadth. It also lists the core workflows that fail its own alpha gate. Four of AutoCAD's five fail: precise drafting, blocks, plotting and a DWG that survives a round trip. Only dimensioning passes.
None of that is on the website. The website says "clean-room reimplementation".
The 3D bit
Here's the line in the feature table that explains my missing humanoid:
| Area | Weight | Parity | Remaining (h) |
|---|---|---|---|
| 3D modelling, visualisation, rendering | 5% | 2% | 300–450 |
Two per cent. The gaps list spells it out:
No 3D: UCS, 3D views, orbit, solids, surfaces, meshes, sections, visual styles, materials, lights, rendering (150 menu items)
The generated parity checklist then lists every one of them as not covered: Box, Sphere, Cylinder, Cone, Torus, Wedge, Pyramid. Extrude, Revolve, Sweep, Loft. Every mesh primitive. 3D Move, 3D Rotate, 3D Mirror, Interference Checking. Top, Front and SW Isometric views. Point lights, spotlights, the Materials Browser. STL export sits in a separate list of missing exchange formats ("ACIS/STL"), at low-to-medium priority.
The 3D milestone itself, M11, is scheduled for "after beta". Beta is estimated at 600–930 agent-hours away.
Five per cent of the time, thirty-one per cent of the menus
3D is a third of AutoCAD's menus but only a twentieth of its weighted score. The number I keep coming back to isn't the 2%. It's the 5% beside it.
CADCraft's readiness figure is weighted by "the share of a typical AutoCAD user's working time". Most AutoCAD users are 2D drafters, so 3D gets 5% of the weight. That's a defensible choice, and the document says so openly.
Look at what it does to the arithmetic, though:
- 3D is 150 of AutoCAD's 491 menu items, nearly a third of the surface area.
- Its 300–450 hours are around a quarter of everything left to build.
- In the features score, 2% of a 5% weight is worth 0.1 points. Finishing 3D completely would add about five.
So the headline barely moves whether 3D exists or not. There's even a separate breadth figure for "2D menus only", which rises to about 65% by leaving those 150 items out.
None of this is dishonest. Every number is labelled, sourced and caveated. But a weighted average describes a typical user, and I wasn't one. For a 2D drafter, CADCraft is perhaps a quarter of the way there. For someone who wants an STL, it's at two per cent.
A week old
Then I found the progress log, which reframed the whole thing:
First version built in one session (≈ 6.9 h elapsed, ≈ 11.5 agent-hours)
That was on 7 October 2026. I installed it on the 13th. CADCraft is a week old, and it was built by AI agents: every estimate in the roadmap is in agent wall-clock hours for a Claude variant called Opus 5.5, which aren't directly comparable to human working hours.
The log also records something I found more interesting than any single number. The first assessment, on 7 October, used percentages "reported by the agents that built each milestone". Three days later the whole thing was re-measured "by depth, not presence". The table below shows what happens when human validation meets AI confidence:
| 7 Oct (self-reported) | 10 Oct (re-measured) | |
|---|---|---|
| Drafting core (M1) | ~85% | ≈ 55% |
| Files (M7) | ~55% | ≈ 30% |
| Remaining work | ≈ 570 h | ≈ 1,300–2,100 h |
| Stage | alpha | pre-alpha |
The agents marked their own homework generously. Once someone checked whether options actually worked, rather than whether a command merely existed, the estimate of remaining work more than doubled.
That's the long tail Claude had warned me about, measured by the project itself. Presence is cheap: around 290 commands in under twelve agent-hours. Depth runs at 0.3–0.7 hours per option or bug fix, and the backlog of those comes to most of the 1,300–2,100 hours left. A 3D solid-modelling kernel, budgeted at 300–450 hours on its own, is roughly thirty times the effort of the entire first version.
Meanwhile, in my own plumbing
Asking CADCraft for its full command list did break something, and it was mine.
The reply was 143 KB of JSON on a single line. meta-mcp read sub-MCP replies with asyncio's readline(), whose default buffer is 64 KiB, and fell over:
Separator is not found, and chunk exceed the limit
The one-line fix was to raise the limit. While Claude was in there, it found two more problems. Each sub-MCP's stderr was a pipe that nothing ever read, so a chatty server would eventually fill it and hang mid-call, silently. And the MCP SDK wasn't pinned, so a fresh uv sync would install 2.x, which removed the FastMCP class meta-mcp is built on.
All three are fixed now. Large results get written to a file, with a path and a preview returned in their place, so a 143 KB command catalogue costs Claude 2 KB of context instead of something like 35,000 tokens. Credit to CADCraft for that one: it found the bugs by being thorough.
What I'll tell my friend
I didn't send the humanoid.
What I'll send instead is a link to ROADMAP.md, with a note: read this before the README. ArtCraft is genuinely remarkable — a pre-alpha AutoCAD clone that opens DWGs, speaks MCP and was built in a week, with engineering documentation more honest than most commercial products manage. The MCP server working first try, in particular, is not nothing.
But the honesty lives in the repo, and the headline lives on the website. "Clean-room reimplementation of AutoCAD" and "3D modelling: 2%" are both true. Only one of them made it into the link my friend sent me.
The 7 October numbers show the cheap part. The 10 October re-measurement shows why the expensive part takes so long — and why the four-million-dollar estimate might still hold.
Martin Harris connects cloud infrastructure to the real world. He asked an AutoCAD clone for a 3D figure and got a very well-documented apology instead.
If you find this blog useful, please consider a donation.