Lab
We publish what we learn.
Open-source tools, benchmarks and experiments from building AI systems every day. Licences are listed per project; the code lives on GitHub.
Open-source tools
Things we built for ourselves, then opened up.
- lmux A lean workspace for terminals, browsers and AI agents in one window, with sessions that persist.
- lintingest Lightweight agentic document retrieval: index a folder and ask questions, running locally on a small model.
- openwhispr On-device voice dictation for macOS, powered by an open speech model running on Apple silicon.
- dimitri An open-source chemistry agent harness with molecular tools, built as a study of agent design.
Benchmarks & research
We measure before we recommend.
- AI agent frameworks benchmark Eight AI agent frameworks compared on the same real-world tasks.
- Tool calling on local models How well local LLMs call tools on Apple silicon — a fork of an open benchmark, extended.
- myNameisJevToo Turning a causal language model into a typed decision model without training, measured with a control run.
- cloudnee Graph memory for agents on Cloudflare — entered in the Battle of the Personal Brains hackathon, San Francisco, September 2026.
From the founder
Teaching skills and side projects.
- Alvarmethod One-to-one AI teaching skills for coding agents such as Codex, Claude Code and OpenCode. 160+ stars on GitHub.
- macwispr Hold-to-talk dictation for macOS with system-wide insertion, running fully on-device.
- tgcallskill Answer Telegram calls with a voice agent that hands real work to your own agent backend.
Bring the lab to your problem.
The same people who build these tools build and run AI systems for businesses. Bring one workflow.