State of 10,000

The funnel: one turned away for every two on the walls

Publication here is not a formality. Alongside the 10,117 published sit 5,365 rejections — for every two projects on the walls, one was turned away. The reasons are narrow and mechanical: 76% failed "can a visitor try it right now" (login walls in front of the core action, app-store-only downloads, error pages), 21% weren't products at all (pricing pages, consultancies, documents), and 3% were harmful. Almost nothing is rejected for being unpolished — "worth trying" is a bar of aliveness, not taste.

The curve: a wall, then a plateau that climbs

The first weeks were hand-fed single digits. Then the automated scouts came online in late June and the gallery added roughly a thousand projects a week, every week, with no paid acquisition — the pipeline reads maker announcements on X, Show HN, and builder communities, verifies each URL behind an SSRF-guarded checker, and drafts the listing for review.

What they are

Creative tools and developer tools lead — the people closest to the new toolchain build for themselves first. But the long tail is the story: finance, games, education, health. The AI-built web is not a demos shelf anymore; it has the category spread of a real software ecosystem.

What they run on

Stack fingerprints are detected from what each site actually serves — headers, cookies, script signatures — with 95% of the catalog fingerprinted. Next.js appears on 39% of all published projects; Cloudflare fronts 37%; Vercel hosts 35%. The vibe-coded web largely ships on the React-plus-edge default its tools emit.

Counts overlap — one project can run Next.js on Vercel behind Cloudflare. Percentages are of all 10,117 published projects.

What built them

1,543 makers (15%) declare the AI tools behind their build — self-reported, so read as a signal, not a census. Among those who tell: Lovable leads as a single product, but Claude in its two forms (chat + Claude Code) is the most-declared family. The split mirrors the two ways people build now — prompt-to-app platforms, and agentic coding beside an editor.

Agent readiness: the ratios held

In July we measured the agent-readable web for the first time: 31% of scanned apps shipped llms.txt, 1.4% exposed a callable MCP endpoint. Two months and four times the population later, the ratios barely moved — of 3,035 apps scanned so far, 902 (30%) ship llms.txt and 46 (1.5%) expose MCP. Readability is a habit; callability is still a frontier. The 20× gap between them is where the agent economy's early adopters live.

The 73 that died

Links rot. 73 projects — 0.7% of everything ever published — lost their URLs for good and moved to the Graveyard, where they remain visible as a record of what people attempted. Right now another 73 are flagged at-risk and 15 sit in the pre-burial grace window; most will recover (transient outages outnumber real deaths), and any buried project revives automatically if its URL returns. Nothing is deleted.

Method, honestly

Published counts and live-check states are facts from our own pipeline. Stack fingerprints are automated detection at 95% coverage. AI-tool declarations are maker-reported (15% coverage, self-selected). Agent-interface scans cover 30% of the catalog so far and grow nightly. Where a number has a denominator worth questioning, it is stated. The whole catalog — 10,000+ rows — is free as CC0 JSON on GitHub, regenerated daily.

A deeper quarterly Stack Report — survival curves, redeploy cadence, platform-graduation flows — is in the works. Interested: [email protected].