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

I'll pull the source thread and HN item so the paragraphs stay factual and specific.Alon, posting as Alegend on the Vintage Computer Federation Forums, reports that after more than 22 years of hunting for a 1980s MATLAB binary to run on a MicroVAX II, he used Claude Code over a couple of afternoons to clean-room re-implement that era of the language. The only original he had ever recovered was the 1982 FORTRAN release, which he had previously compiled and run on hardware. What he actually wanted was the later pre-GUI, command-line MATLAB 3.x class of product: user functions, script files, and a signal-processing toolbox. He had even contacted Cleve Moler years earlier; Moler was interested but never followed up, and later died. Moler’s 2018 MathWorks blog post, which listed keywords, operators, and functions from the oldest PC-MATLAB disc he still held (version 1.3), became the functional target. The result is VAXLAB, a clone Alon describes as close enough to version 3.5 to be usable, with m-file execution so toolboxes can grow by writing functions rather than shipping a frozen binary.

Technically, Claude Code drove almost the whole stack under minimal human steering. Alon supplied the compiler (Compaq C V6.4-005 on VAX/VMS), the feature list, and answers to design questions; the model smoke-tested the toolchain, staged work into milestones, and built a full m-file test suite for every feature. Compiles ran on a real MicroVAX II, though SIMH would have been valid hardware too. The shipped CLI covers matrix arithmetic and linear algebra (including inv, eig, svd, lu, qr, chol), control flow and function definitions, signal routines (fft, filter, Hamming/Hanning/Blackman windows), file I/O, and graphics with ASCII plus REGIS/SIXEL backends Alon requested. Claude’s own build recap describes C89-only VMS constraints, matrix literal and operator semantics, a process-wide global symbol table with per-scope redirects, and general non-symmetric eig via unshifted QR (modified Gram-Schmidt), with hand-traced iterations on a test matrix and hard errors when complex eigenvalues cannot be represented. Live demos show matrix inverse recovering the identity and sinc-style plots on the terminal. Alon notes limited testing has not yet found bugs, but expects some remain.

For engineers and builders, the point is not nostalgia theater. It is a concrete human-in-the-loop pattern: a domain expert with a frozen public surface (the 1.3 feature list), real target hardware, and a closed compile-run-report loop can treat an agent as the implementer while staying a supervisor who only ratifies milestones and reports failures. Alon states he never needed to know parser structure or Gram-Schmidt QR to ship a working product. That compresses work that he had long considered unrealistic for a solo effort into a few afternoons, while still requiring someone who can run the binary on VAX, read compiler diagnostics, and judge numerical and language correctness.

Market and competitive context is the dead-software recovery problem that archival searching and vendor outreach never closed. Original discs and licensed VMS MATLAB are scarce; another forum member still holds MATLAB 4.2c (May 10 1995) with Control, Signal, Stats, and Neural Network toolboxes, but that is a later, richer product and is tied to MAC-address licensing. Clean-room recreation from a published feature inventory sidesteps missing media without reverse-engineering a proprietary binary. On Hacker News the same story showed only 1 point and 0 comments at capture, so wider developer attention is still thin relative to the technical claim. Adjacent tooling discussions on the thread (Digital logic sim, MATRIXx, other abandoned packages) treat VAXLAB as a proof that agent-driven reimplementation can apply beyond this one language.

Practical takeaway: if you need extinct software, start with the tightest public behavior contract you can find, pin the real compiler and OS, and force the agent through a hardware-validated milestone loop rather than a single big prompt. Watch for published prompt logs and source drops of VAXLAB, independent runs of its m-file suite on MicroVAX or SIMH, and whether the same method extends to later MATLAB 4.x feature sets or other lost numerical packages. Token cost and residual numerical bugs remain the main risks Alon already flags when the target grows richer.