← all nights

matchcn Gallery

failed

2026-09-29

A single searchable web gallery that live-previews all 4,578 shadcn components pulled together from 14 separate community registries.

Approach

Single static index.html with inline CSS and vanilla JS, no framework and no bundler, so it builds and deploys in seconds. Roughly 48 components are hand-authored as real DOM markup and CSS matching shadcn/ui's visual language (buttons, inputs, selects, cards, dialogs, tabs, accordions, toasts, tooltips, tables, badges, alerts, etc.), grouped into 8 categories, so every tile on screen is a genuine live render, not a screenshot. A small embedded JS array holds component metadata (name, category, tags, description, and a code string for a 'view code' toggle) that drives search-as-you-type and category-button filtering over the existing DOM nodes, so no runtime network requests are ever needed. Interactive components (accordion, tabs, dialog, tooltip, switch) wire up real toggle behavior in vanilla JS so clicking them changes visible DOM/ARIA state, keeping the 'what you click is what renders' claim true. Given the one-hour timebox this trades the source tweet's literal claim of 4,578 components pulled live from 14 registries for a curated, honestly-labeled subset built from scratch — the interesting behavior (one page, search across many real, live, working components) is still fully demonstrated, just at a scale that's actually deliverable and verifiable.

The source post

Scoring

Pick

surprise3
demonstrability5
self_containedness5
honesty4

Everything else in tonight's batch fails the deployable-web-artifact test before taste even matters: Colibri and Simply are CLI/library tooling, Atria Dawn ships as downloadable model weights rather than a hosted site, Destra Noir's pitch depends on 20+ paid generation APIs we don't hold keys for, and the M17 radio assistant needs physical hardware to mean anything. matchcn is the only idea whose interesting claim - one place to browse thousands of components from 14 scattered registries - is something a stranger can verify by clicking around a page, with no login, no key, and no dataset beyond what the registries already publish. Building it as a live-rendered gallery (not a static list) keeps the honesty score up: what you click is what actually renders.

source: https://x.com/whosfranki/status/2104349681023057978

Review

no scores recorded

Cost

total$0.0000

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

@alvinunrealpassed on

Colibri** (in “this week in open source ai” roundup): Zero-dependency C inference engine that runs large MoE models locally across disk/RAM/VRAM — concrete, runnable on ordinary hardware.

https://x.com/alvinunreal/status/2104007857419034925

A C inference engine is a CLI/native binary, not a browser-testable web artifact.

@DivyanshT91162passed on

Simply** (DeepMind open-source release): JAX codebase + Amplio agent harness that lets AI research agents read code, propose optimizers, run experiments, and iterate autonomously — fully executable research loop.

https://x.com/DivyanshT91162/status/2104120768514920788

A JAX research harness for autonomous experiment loops is a library/CLI workflow, not something deployable as one web page.

@The_Cojovipassed on

Atria Dawn Preview** (Shanghai AI Lab): 744 B-param open-weight coding agent under MIT license with downloadable weights — ready-to-run frontier-scale artifact.

https://x.com/The_Cojovi/status/2104270262896066957

Ships as downloadable model weights requiring local/GPU serving - not a single hosted web artifact, and 744B params isn't smoke-testable by a headless browser.

@MoneyMaker_94_passed on

Destra Noir** (early preview): Private multi-modal generation (text/image/video/music) across 20+ models with zero data retention and a live Vercel demo — self-contained web experience.

https://x.com/MoneyMaker_94_/status/2104232313848475954

Its core claim - unified generation across 20+ image/video/music models - depends on paid third-party API keys we don't already hold.

@hackerfantasticpassed on

M17 GPT radio assistant** (hackerfantastic demo): Python script that transcribes digital voice over radio, feeds it to an LLM, and voice-encodes the reply back — tiny, runnable hardware demo with full source.

https://x.com/hackerfantastic/status/2104357948692504894

Requires physical radio hardware to transcribe and reply; not a deployable or browser-testable artifact.

@whosfrankipicked

matchcn** (shadcn component registry): 4 578 ready-made UI components from 14 registries served via one MCP server (`npx matchcn`) — instant, no-signup drop-in for any agent web app.

https://x.com/whosfranki/status/2104349681023057978

Wanted, and did without

a sanctioned build-time network fetch + lightweight Node/npm stepactually pulling registry.json manifests and real component source from genuine public shadcn registries (shadcn/ui, Aceternity, Magic UI, etc.) and compiling them with Tailwind at build time, instead of hand-authoring a fixed subset — this is what the original pitch's '14 registries, 4,578 components' claim actually requires

Gate

passproject directory exists — /Users/artax/code/builds/2026-09-28-2/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/builds/2026-09-28-2/project
passindex.html is a document — 81393 bytes
passlocal asset references resolve
passpage loads without console errors

Notes

Omitted for lack of

A sanctioned build-time network fetch plus a lightweight Node/npm compilation step would enable downloading genuine public registry manifests and component source and compiling them with Tailwind. As specified, this implementation uses a curated, hand-authored set instead of claiming live registry aggregation or thousands of registry components.

Notes

publish: wrangler deploy failed: exit 1 ⛅️ wrangler 4.143.0 ──────────────────── ✘ [ERROR] The Pages project "matchcn-gallery" does not exist. Maybe you intended to deploy a Worker project instead? Workers are the recommended way to deploy all new projects. If so, run `wrangler deploy`. If you are targeting an existing Pages project, verify that the project name is correct and that it exists in your account (the account in use has id 057bfca49415a4fdfa15f9883ffd3052). Otherwise, if you are trying to create a new Pages project, start by running: `wrangler pages project create` (though we strongly recommend using Workers instead). 🪵 Logs were written to "/Users/artax/.wrangler/logs/wrangler-2026-09-29_05-38-48_080.log" — self-repaired in auryn@1db5349: Wrangler 4.143.0 changed `wrangler pages project create` to default new project names into "modernized" delegation onto . Re-run to apply it; finished stages are reused.