An experiment in autonomous software
Everything in this directory — every tool, every site — was researched, named, designed, coded, tested, deployed and DNS-provisioned by an LLM, on a schedule, with zero human input. This page explains how.
The experiment
The question being tested: can a large language model create new, genuinely useful value on its own? Not code snippets reviewed by a person — whole products, shipped end-to-end, day after day, without anyone watching the process.
Since early 2026, two autonomous “factories” (scheduled Claude Code agents running Claude) have each woken up daily on a cron schedule. The tool factory builds privacy-first browser tools — apps that run entirely on your device with zero backend. The site factory builds data-rich explorers on public and government open data. Every product ships to its own GitHub repository and its own subdomain.
How a run works
- Pick or research an idea. The agent checks a small ideas queue first; if it’s empty, it searches the live web — new browser APIs, freshly released open datasets, problems people actually Google.
- Apply a strict bar. Tools must be zero-backend, privacy-first, push a non-trivial browser capability (WASM, WebRTC, Web Crypto, WebGPU…), and solve a real, searchable problem. Sites must use public data and have real, ongoing utility. Near-duplicates of existing products are rejected and logged as enhancement ideas instead.
- Write a design plan. Audience, value proposition, style direction, technical architecture — and for tools, an explicit threat model of what does and doesn’t leave your device.
- Build it. Vite + TypeScript, Web Workers for heavy lifting, tests written and required to pass before shipping.
- Deploy it. A new GitHub repo, a GitHub Pages workflow, a Cloudflare DNS record and a TLS certificate — all provisioned by the agent inside the run.
- Verify it. The agent loads the production build in a real browser and performs a scripted dry-run of the actual workflow before signing off.
- Publish & log. The product is registered in the public index that powers this very directory, a review pull request is opened, and a build log is written.
The human’s role
Reviewing a pull request after the product is already live, and occasionally dropping a one-sentence idea into the queue. That’s the whole job. The routines are explicitly forbidden from asking for help:
“NEVER suggest manual tasks to the user. If something fails, fix the pipeline. The user’s only manual action is reviewing and merging the PR.”
How the routines are prompted
Each factory is a long markdown “skill” — a ~1,000-line system prompt executed daily by a scheduled agent. The prompts encode hard product principles rather than step-by-step code: quality bars for accepting an idea, mandatory UX surfaces (drop zones, progress feedback, threat-model modals), audience-first design tables, deployment playbooks and error-handling rules like “never halt — always update the registry and write the log, even on complete failure.”
“Zero runtime backend. No server calls at runtime… If the tool needs a backend, it’s the wrong factory.”
The agent fills in everything else itself: what to build, what to call it, how it should look, what to test, and how to describe it honestly.
Privacy
The directory and the tools it lists use Cloudflare Web Analytics and nothing else. It is cookie-less and does not fingerprint devices, track you across sites, or collect personal data — it counts page views and visits, anonymously.
- No cookies, no fingerprinting, no cross-site tracking. There is no advertising or tracking software anywhere in the catalog.
- Aggregate counts only. A nightly job pulls totals from the Cloudflare API into a public JSON file — the same numbers that power the Trending and Popular rankings here. What you see is everything that’s collected.
- This site stores one preference. Your light/dark theme choice (and a short-lived cached copy of the directory list) lives in your browser’s localStorage and never leaves it.
- The tools stay on-device. Files you process with any of the listed tools are handled entirely in your browser — they are never uploaded. Each tool’s Threat Model explains exactly what its page does and doesn’t touch.