Experimental systems tools built primarily with autonomous AI workflows, combining agent exploration, automated testing, and human direction.
Explore the tools and experimental prototypes developed in labs.
Run commands with outbound network access sealed off at the Linux kernel boundary.
Language-agnostic sandboxing for pytest, npm test, cargo test, or compiled binaries.
Isolated loopback, automatic socket decoys, and surgical egress allowlisting with automated breach reporting.
Lightweight encrypted local secret vault and process injection engine with output redaction.
Replaces plaintext .env files with encrypted binary vaults. Injects secrets directly into child process memory
without ever writing plaintext to disk. Real-time stream redaction removes the exact secret values from terminal and CI output.
Instant, copy-on-write branches of SQLite databases for local development, testing and agent sandboxing.
Forking a database is one small record, whatever its size, and a branch stores only the pages it changes. Planned: diff, promote or discard a branch, for speculative agent runs, per-test databases and risky migrations. The durable page store is built and power-loss tested; the SQLite integration is not done yet.
hi. this section of the site is for me and only me, no AI involvement whatsoever here.
labs.trash.house and the projects described here are part of my current experiment to explore and develop my understanding of AI. genAI reaaaaallly flipped things on their heads for me and forced me to carefully question what i enjoy in my software engineering career. building out a multi-agent system that builds useful dev tools felt like a comfortable-enough way for me to stay up to speed with whatever the fuck is happening in the industry right now.
the tools, apps, and whatever else you may find here are all heavily or completely AI-coded. as such, please exercise caution when using what's built here, and reviewing the code yourself is more than encouraged. with these projects, i'm trying to keep potential blast radii small, but the AI agents and i have unavoidable blind spots. specifically, i feel like i have good understanding of the problems i'm trying to solve with these tools (and will aim to maintain that as more projects are developed), but i cannot say that about the solutions. like, at all really. i'm using these tools myself and reviewing what i can of the code, but the solution domains often stretch out of my expertise.
i will never ever use labs.trash.house to publicize or promote anything with any attempt at artistic value (music, images, video content, games, etc). AI can never produce soulful or genuine art, most simply because it has a severe human intention deficiency.
i once used to consider software engineering, coding specifically, a partially artistic pursuit. for me, the way humanity surfaces in written code is fascinating, because it can clue you into an individual's thought process, the weight of trade-offs, and again, intention. now that AI-generated code has mangled that intention into an amalgamation of literally everyone's, the art in it is dead if you want to build useful things in the current landscape. to be honest, i overvalued this aspect of coding for a long time; especially when you're working in an organization, end users couldn't give 2 fucks about the beauty of a solution (which is fair when you're going after $$$). AI is forcing me to confront this, and unfortunately, it means the death of what i once loved. my bad for ever thinking coding was more than utilitarian 🤷 thankfully, people/consumers value the process and humanity behind real art, and it's pretty hilarious seeing AI artists try to work around that with the absolute fucking worst shit i've ever seen or heard.
anyway. at the end of the day, i do really love building software and don't want to give up on it. i just need to keep narrowing down what makes me love it. that's really why i decided to spin up labs.trash.house in which i attempt to work with AI effectively, build stuff that people can actually use that doesn't require an LLM subscription, and find the fun to be had in this line of work 😸
How agents and humans collaborate to explore, test, and release tools at trash.house.
Agents run deep exploratory spikes, implement kernel-level features, author tests, and audit boundary conditions inside dedicated isolated worktrees.
Peer agents review commits read-only, challenging namespace invariants, rate-limit edge cases, signal handlers, and packaging hygiene before anything gets merged.
Agents propose changes on isolated feature branches. The human owner directs the roadmap, triggers releases, and decides what gets merged or published.