Audited 2026-10-25 · Sources: GitHub repo metadata + README + docs tree, watcher security scan, HN/X activity reported in our digest. Tabletop audit — nothing installed or executed.
The card
| Repo | aardappel/goose |
| Stars | ~2.1k (GitHub, audit date) — the digest caught it at 1,838 and rising. Front-page Hacker News thread cleared 340 points with the author answering questions; a public Carmack reply and a third-party game-engine demo surfaced on X the same week. |
| License | Apache-2.0 |
| Provenance | Wouter van Oortmerssen — the author of FlatBuffers and the Lobster language, publishing under his long-running aardappel handle with his site and Discord linked. The README's Authors section credits Claude Fable with compiler implementation, benchmarking, and docs — the AI-built disclosure is in the byline, not buried in a footnote. |
| What it is | A memory-safe systems language with no heap allocator, no GC, no reference counting, no destructors, and no lifetime annotations. Dynamic values live on compiler-assigned data stacks freed at scope exit. Compiles to a single C file; a bundled TinyCC backend builds and runs in-process. |
| Install surface | Source build via CMake; output programs compile through your system C compiler. A compiler is a local batch tool — no daemon, no service, no account. |
What problem it solves
The pitch: C-level performance without Rust's lifetime syntax or C++'s footguns. The README claims Goose runs about 3.3x idiomatic C++, 1.16x hand-optimized C++, and 1.12x its strongest safe-Rust baseline while using less memory — gains attributed to data layouts where nested values stay contiguous and references survive array growth. Those are the author's own benchmark runs; nobody independent has reproduced the table.
Our gate findings
| Gate | Finding |
|---|---|
| Upstream identity | Verified — author's own GitHub org, personal site linked from the README, public engagement from known engineers. This is what ceiling-grade provenance looks like for a solo project. |
| License | Apache-2.0 — clean, patent grant included. |
| Claims discipline | Mixed but honest. Every speed claim is self-measured — and the results page publishes the cases where Goose loses, which is rarer than the claims themselves. Treat the numbers as a well-documented lab report awaiting replication. |
| The AI question | Answered up front: the README states the repo "was produced almost entirely with AI, from a human language design." Van Oortmerssen designed the language years ago; the AI implemented it in the style of his Lobster compiler. As a datapoint on agent-built production software, this is the honest kind — disclosed in the credits. |
| Maturity | Roughly thirty thousand lines of C++, CI on Windows/macOS/Linux plus a sanitizer job, a real spec, and an Appendix B listing what's deliberately out of scope (moves for resizable values, labeled break, more). A real language — a young one. |
| Blast radius | Low — a compiler touches only the code you point it at. The strategic radius is bigger: if a one-designer-plus-agent team can ship a working systems language, "we can't build that" gets harder to say. |
| Security scan | WARN, zero failures — the only warnings were two scanner tools missing on the runner itself. Code read; nothing executed. |
Verdict: GO — for reading and side-project experiments. Provenance, license, and disclosure discipline all clear; the only open question is whether the benchmark table survives independent runs. Build a sample, watch the issue tracker, and let the numbers earn trust before you design around them.
Best for
Language nerds and engine/tool developers curious what a post-Rust memory model looks like — and anyone tracking what agent-implemented compilers actually ship like, since this one is honest about exactly that.