This is where the site tries things. Each experiment lives in its own corner and touches nothing else, so ideas are cheap to try. The ones that prove useful get pulled into every page; the rest stay here. Two rules carry over from the site: everything runs in your browser with no tracking, and every tool works for you and for the AI agent working for you.
Pick any tool and get it in the form you need: a link for a person, an embed for a website, commands for a terminal or MCP client, or one prompt you paste into a coding agent that tells it how to put the tool into your own project.
Distribution where the tool travels with its own instructions. An agent can integrate a tool it was handed in one step instead of reverse-engineering a web page.
Every tool gets its own cover art, drawn from its name by a few lines of code: same name, same picture, no image files, no AI bill. You can see them in the handoff experiment.
A catalogue of 1,000+ text cards is hard to scan. A stable picture per tool gives the eye a landmark, and the same function can draw social cards.
Press / or ⌘K anywhere to jump to any tool, category, guide or post, and run "hand to agent" from the same box.
Navigation for people who know what they want, and one place an agent can drive with the keyboard.
A front door that starts from "what are you trying to do?" and answers with the two or three tools that fit, matched in the browser against catalog.json.
Most visitors arrive with a problem, not a tool name.
A plain-Markdown copy of each page beside it (`index.md`), linked with rel="alternate", carrying the inputs, the formula and the disclaimer.
Assistants read text better than forms. llms.txt points at the site; twins would let it point at every page.
A reading mode on any result that shows the arithmetic step by step, in words, with the inputs that moved it most.
A number people can follow is a number they trust and come back to.
Short walkthroughs recorded from the live pages by the browser checks that already drive them (Playwright), captioned from the page copy. No camera, no paid generator.
Short-form video is a discovery channel, and the site already has a robot that clicks through every tool.
Each research project in `research/` as a short write-up, a 60-second short (its key chart, animated), a long-form walkthrough, and a podcast episode, all generated from the project’s own LOG.md, data and charts, so nothing is said that the research did not find.
Research people can build on has to reach them first; the findings, figures and logs are already committed, and a pipeline that renders them costs nothing per episode.
The posts in `writing/` and the research write-ups as a podcast feed, voiced in the listener’s own browser (Web Speech) at first, and by a model on the owner’s key only if the owner switches that on.
The engineering writing already exists; audio is a second door into it at no marginal cost.
Ideas are rows in EXPERIMENTS. To try one, build it under lab/<id>/ and flip its status to live.