Hobbyist on one Linux box, with a few too many reverse-engineering tools. Claude Code does the hands-on work. It runs the tests, does the reverse engineering and drafts the write-ups. I pick what it looks at, steer it, and read every result before it goes up.
Mostly I point it at other people's projects, so most repos here are forks. One is my own.
a10e-firmware (GPL-3.0): Why my locked Galaxy A10e won't take LineageOS. It's my phone, and I asked the questions. Claude read the stock firmware. Scripts and a full write-up, plus a short version on REA's Discussions.
| Project | What it is | What came of it |
|---|---|---|
| REA (MIT) | Reverse engineering with AI agents | Linux tests and code reviews on 216 PRs. 35 issues filed, 24 of them bugs. 6 of my own PRs merged, like website dark mode and an Electron launch fix |
| rsleigh (Apache-2.0) | A decompiler workbench in Rust | Made output the same every run. Recovered two ls jump tables. 2 bug reports |
| kuna (Apache-2.0) | A decompiler built for agents, in Rust | Speed and memory tests on big binaries. 2 bug reports |
Counts as of 11 Nov 2026.
- audio.cpp (Apache-2.0): Speech and audio models in pure C++. Forked to test on Linux.
- abtop (MIT): A monitor for AI coding agents.
- lot-map (MIT): A live map of what Claude Code is doing.
Nothing of mine is in them yet.
Claude reads the diff first. Then it gets a real run on my box: the build, the tests, and the thing the PR says it fixes. The result goes on the PR as a comment or a review: what ran, with exit codes, what didn't, and the box it ran on. Every post ends with a footer saying what the AI did and that I OK'd it.
By 8 Nov 2026, 95 PRs Claude tested or wrote had merged on REA.
Many of the reviews check what the review bots claimed, on a real box. Three of the 216 PRs that caught something before merge:
- 34 of 35 bot findings didn't reproduce. Claude found a new bug instead: an upper-case Accept header broke the live session. It was fixed within half an hour.
- A bot said a new default would break REA's own check scripts. Claude replayed the scripts' checks, and they failed. The scripts were fixed before merge.
- A crafted ELF file hung the parser for up to 112 seconds. It still said the parse was complete. Fixed and retested.
- REA on Fedora-based Linux: Setup notes, gotchas and my test queue.
- Benchmarks: REA's startup time on Node 22, 24 and 26, plus five local models compared.
- My local model setup: Free local models reading REA's output on one 12 GB GPU.
- Local models can drive REA over MCP: With about 10 tools showing, the right answer 3 of 3 times on a 27B model and 4 of 5 on a 35B.
- kuna, rsleigh and Ghidra on the same three files: First speed results, side by side.
My test bench: Nobara Linux 44 (Fedora-based), Ryzen 9 5900X, 32 GB RAM, Radeon RX 6700 XT 12 GB. Node 22, 24 and 26, Ghidra 12.1.4 (Apache-2.0), and local models in llama.cpp (MIT).