Using Claude to re-create extinct software – Matlab from the 1980's

I'll pull the forum thread and HN item so the paragraphs stick to stated facts only.Alon, posting on the Vintage Computer Federation forums, reported a clean-room re-implementation of 1980s-era MATLAB aimed at a real MicroVAX II after more than 22 years of searching for a working binary from that era. The only historical source he had run successfully was the 1982 FORTRAN release; the target he wanted was roughly MATLAB 3.x — pre-GUI, command-line only, with user functions, script files, and a signal-processing toolbox. After contacting Cleve Moler years earlier (Moler later said he could not locate a disc of the first PC-MATLAB and had only version 1.3’s feature list) and finding a full rewrite impractical by hand, Alon spent a couple of afternoons with Claude Code and produced VAXLAB, dated 15 Jul 2026 in the demo session.

Technically the build was constrained to the host toolchain and machine, not a modern laptop simulation first. Claude smoke-tested Compaq C V6.4-005 (DEC C / C89 on VAX/VMS), wrote BUILD.COM and small probes for sizeof and argv case behavior, then staged the language: matrix literals and operators, builtins, control flow, m-file execution, file I/O, and linear algebra including inv, det, eig (Jacobi for symmetric cases, unshifted QR / modified Gram-Schmidt for general eig), plus fft, filter, and classic window functions. Graphics cover ASCII plots by default and REGIS/SIXEL when requested. Demos show matrix inverse checks (a * inv(a) yielding the identity), sinc-style plots, and a large suite of m-file tests (TEST*.M, TLINALG.M, TDSP.M, and others). Alon said limited testing found no failures so far, while still expecting hidden bugs.

For engineers and builders the useful pattern is not “AI wrote a toy,” but human-as-hardware-oracle plus agent-as-implementer. Alon supplied the feature list, the compiler, and pass/fail reports from real MicroVAX builds; Claude scoped milestones, chose designs (for example preferring argv over LIB$GET_FOREIGN for case-sensitive names), redesigned a global symbol table with is_global redirects, and hand-traced numerical steps such as QR iterations on [4 1; 2 3]. Someone who does not know parsers or Gram-Schmidt can still ship a multi-module C interpreter if they can run the binary and answer milestone questions. That compresses work that used to mean months of language and numerical library craft into a few afternoon sessions of tight feedback.

In market and tooling context this sits beside other agentic coding stacks people already compare on HN-style threads — Claude Code here, with Codex often in the same conversation as a rival agent — applied to a niche that commercial MathWorks product lines no longer serve: extinct workstation software on DEC hardware. Preservation communities usually depend on disk images, backups, or rare licensed installs (one reply still holds MATLAB 4.2c from May 1995 on a VAXstation, MAC-tied licensing and all). Clean-room clones from published keyword/function lists sidestep missing media when the behavior surface is documented well enough, without claiming bit-for-bit compatibility with a lost binary.

What to watch next is whether VAXLAB’s sources and prompt history get published for independent compile-and-test on MicroVAX or SIMH, how far m-file toolboxes stretch beyond the built-in signal-processing set, and whether the same human-in-the-loop hardware loop works for other dead packages people still want on VMS and similar platforms. Token cost and numerical edge cases (complex eigenvalues, format edge cases, buffer sizes around VAX D-float range) are the practical limits Alon’s own write-up already flags.