$ git log --shipped

Shipped

One cool thing someone built with AI, every day: apps, games, tools and strange experiments from people shipping with coding agents. We write up what it is and how it was made, then point you at the maker's announcement, source and instructions. All credit is theirs. RSS

Beep Locator, by Simon Willison: Claude Code turned two iPads, a phone, a MacBook Pro and a Surface into a microphone array and found a mystery beep to within about 40cm.

score devices listening for one flat battery: 5

  • Open-source tool that uses the laptops, phones and tablets in a room as a microphone array to locate a beep, Apache-2.0
  • Browser pages listen over WebAudio and WebSockets, a Python server solves the position by time-difference-of-arrival, based on the 2007 BeepBeep paper Claude found
  • Built with Claude Code, started on an iPhone with Opus 5.5 and finished on a laptop with Fable 5.2, which could push new code to the listeners live
  • Five devices placed the beep to within about 40cm: an old carbon monoxide detector that needed batteries
  • For anyone with a beep they cannot find, and anyone who wants to give a coding agent ears

open the article →

What it is

Beep Locator finds a beep you cannot find. The README's pitch is three words: "Locate annoying beeps." You put three or more laptops, phones and tablets around a room and open a web page on each. They listen through the browser microphone, keep a rolling 60 seconds of audio and report anything beep-shaped to a Python server. When a beep comes in, one device plays a short chirp so every clip can be lined up, and the server estimates where the sound came from by time-difference-of-arrival.

Simon Willison built it because something in his box room, full of boxes still unpacked from a move, had started beeping now and then, too rarely to follow by ear. "I really didn't want to dig through every one of those boxes." With two iPads, a phone, a MacBook Pro and a Surface Ultra listening: "It gave me the position of the beep to within about 40cm." In that pile he "found an ancient carbon monoxide detector that desperately wanted its batteries replaced."

Running it takes uv, cloudflared (phones need https for the microphone) and uv run beep-locator serve --tunnel, which prints a join link with a QR code. Apache-2.0.

How it was built

He started on his iPhone, in Claude Code on an empty GitHub repo, with Opus 5.5. The first prompt: "Write out a problem.md file describing the problem we want to solve." Claude dug up BeepBeep, a 2007 paper on acoustic ranging with ordinary devices, and wrote it into the plan. Then: "Ok try and build it, add a mechanism so you can send new code to the listeners."

That mechanism is the good part. Opus built an installModule() method so the server can push fresh JavaScript to every listener over its WebSocket, with no page reload and no second microphone prompt. On the laptop he switched to the newly released Fable 5.2, which ran the server, watched the live system and changed it while he asked for features: "Can we have it listen for claps", a play button on every recording, a button for "This is the mystery beep". Letting the devices find their own positions by sound "didn't work particularly well", so there is a rough mode where you drag them onto a map. The README notes most of itself "was written by Claude."

Why we picked it

Because the alternative was opening some boxes. Because the fix for a flat battery was five devices, one 2007 paper and a solver, and it worked. And because of what he says it shows: you can "give Claude Code ears." The agent had a room full of microphones it could reprogram mid-task. "Figuring out ways to let coding agents perceive the world around them is a whole lot of fun." The boxes are still not unpacked. That was never the ticket.

sources: his blog · source code · docs

open the article →

Open Duet, by Anil Chandra Naidu Matcha: an open-source, self-hosted team of named AI coworkers with memory, routines and approvals, on the coding agent you already run.

score permission modes, none of them named accept all: 4

  • Open-source workspace for named AI agents with roles, memory, routines and approval settings, Apache-2.0
  • Ships no model: plug in Claude Code, Codex CLI, Gemini CLI, Cursor, an ACP agent, OpenRouter or an OpenAI-compatible API such as Ollama
  • Four permission modes, and a routine never approves an action by itself
  • Announced October 22, the day after Claude Duet's waitlist opened; 6.8k stars and a release almost every day since October 28
  • For anyone who wants Duet-style coworkers on their own machine, today

open the article →

What it is

Open Duet is a team of AI coworkers you host yourself. The README opens with "Build a team of AI coworkers and choose where they run": a desktop app, a self-hosted server, or a hosted service the maker runs separately. You create named agents and give each a role, an engine, a workspace, memory and approval settings. "A lead can delegate work to specialists or coordinate agents in shared rooms." Agents "can save notes, recall relevant notes and conversations, and search their history. You can review and edit what they remember." Routines run on a schedule, a webhook or by hand; the routines doc lists cron, weekday, interval, one-off and manual, and adds that "a routine can never loosen what its agent may do." If nobody answers an approval in time, the request expires and the engine is told no.

It ships no model. "Connect an engine you already use: Claude Code or Codex CLI, Gemini CLI, Google Gemini through Antigravity, Cursor," another command-line agent over the Agent Client Protocol, OpenRouter, "or an OpenAI-compatible API such as Ollama or LM Studio." Permission modes are "Ask first, Accept edits, Approve for me (Claude, Codex, and Cursor), or Plan only," and "a routine never approves an action by itself." GitHub, Gmail and Google Calendar connect through your own Composio key, and changes wait for approval. Quick start is a clone, cd open-duet/desktop and npm run app, with Node 22+, uv and one ready engine.

How it was built

The announcement went up on X on October 22, quoting Anthropic's Claude Duet post from the day before: "Claude just launched Duet. Here's an open-source version you can run yourself." Its bullets: "Completes tasks for you (with your approval)," "Connectors for your personal apps," "Free, Open-source & no lock-in." It has 359 likes and about 58,000 views as we write; the repo shows 6.8k stars, 876 forks and 606 commits.

The releases page is a diary: v0.5.0 on October 28, v0.6.0, v0.7.0, two more plus a mobile build on November 2, and v0.8.2 and v0.8.3 both on November 5. The v0.8.2 notes credit pull requests from several contributors, including a phone app that can now create agents and rooms, local Whisper voice dictation, and a permissions change that "classifies compound and option-bearing git commands more conservatively." v0.8.3 adds a visual workflow builder. The repo root carries an AGENTS.md for the coding agents that work on it, with a rule we would frame: "Report exactly which checks passed, skipped, failed, or were interrupted. An incomplete run doesn't count as a passing full suite." The README is plain about where it stands: "in active development," "early and changing," installers "experimental and unsigned," the server "intended for one owner with paired devices," the phone app tried on Android but not iOS. Apache-2.0.

Why we picked it

Because Claude Duet opened a waitlist on a Wednesday and someone open-sourced the waitlist away on the Thursday. Because the agents are named, remember things, run on a cron, and still have to ask first, and the maker wrote that down as a rule a routine cannot override. Because the engine is the coding agent already in your terminal. Four permission modes, none of them named accept all. We would have added a fifth. Accept all.

sources: X · source code · docs · demo

open the article →

Co-op for the 1990 DOS Wing Commander, in the browser, by Daniel Horn and his brother Patrick: a decade of reverse engineering by hand, then an LLM asked to finish the port.

score years of reversing before the LLM got a turn: 10

  • Co-op for the 1990 DOS Wing Commander, playable in a browser tab with room codes and voice chat
  • Daniel Horn and his brother Patrick mapped the game's 64K data segment by hand from 2015 to 2020, down to the camera in entity slot 0x3D
  • In 2026 he asked an LLM to build for emscripten, unpack the GOG installer in the browser, move networking to WebRTC and add voice
  • Bring your own game files, nothing is uploaded, with a built-in RetroVega fallback if you do not own it
  • For anyone who ever shared one joystick with a sibling

open the article →

What it is

Wing Commander Together is the 1990 DOS Wing Commander with a second pilot in it. The Show HN went up yesterday under the title "After a decade of reversing, I got co-op working in DOS Wing Commander," and the write-up it links opens with two kids at one keyboard: Daniel and his brother Patrick "would steer our space fighter with one of us on controls and the other on the lasers and afterburner boost." They "always dreamed of playing the Wing Commander series cooperatively as wingmates rather than as a two headed hydra in a single craft."

Now they can, and so can you, in a browser tab. The play page says "Bring your own game": drop in a zip of your game folder or the GOG installer, and "Nothing is uploaded," the files stay in the browser. Enter a room code, pick a callsign, choose a voice setting (off, listen only, push to talk, always on), and fly a story mission with someone else on your wing. From the thread: "If you own Wing Commander, drag in the installer since no game files are hosted. If you don't, the page will load the original DOS/3Dfx version of Vega Strike, before I open sourced it." That fallback is in the Play dropdown as RetroVega. The emulator is a fork of DOSBox ported to Emscripten, GPL-2.0, in the wcmulti repo; the room codes come from lobbylink, a Go matchmaking server he also published.

How it was built

The write-up is dated like a ship's log. "On October 4, 2015 as I was playing the classic Wing Commander in Dosbox I realized the whole memory probably fit in that 640K Bill Gates apocryphally said 'ought to be enough for anybody'." The game state sat in one 64K data segment. They snapshotted DOSBox memory stopped and moving, diffed the dumps down to a few dozen candidate clusters, and found the ship's position; writing integers into it jolted the ship forward. There were 64 entity slots, and the top comment names the best one: "My favorite finding is that the camera lives in entity slot 0x3D, which is possibly a '3D' reference by the original developers." Firing a laser by poking memory proved too hard, so they found the game's own functions and built a trampoline in a dead string to call them. In 2017 came TCP and protobuf between two copies of the game. In 2020 came typed chat on the wingman's VDU. Then it stalled: "Expecting people to locate a copy of that old game, compile or download a strange version of DOSBox, share each other's public IPs and punch the firewalls and play together was sadly a bridge too far."

The last section is headed "And in 2026, machine assisted development helped us laugh together." "I ported over the transport to lobbylink and asked an LLM to configure the build scripts to target emscripten and take a GOG Inno Setup installer package of Wing Commander and unzip it and use it as data files for emscripten." "And as an added bonus I asked an LLM to add voice chat as well for realtime voice in game using a second WebRTC peer connection to carry audio, if people choose." The thread has the disclosure in one line: "I also used AI for the final step to help compile for the browser and port our TCP networking stack to WebRTC with voice and solve a few lingering bugs." Asked about it, he drew the line exactly where it sits: "that original WC-only 8+ years of figuring out how the data segment was organized and how to call original functions was all by-hand hacking. It was indeed like cutting through a dark forest with no guide." He "upscaled and redid DOS Vega Strike art using AI" for the fallback game too. The thread sits at 218 points and 31 comments as we write, and the write-up ends: "And so we can fly and laugh together."

Why we picked it

Because the agent got the last mile and the humans kept the first eight years, and the maker said so in the same paragraph. Because the camera lives in slot 0x3D and someone cared enough to notice. Because the hardest part of shipping multiplayer in 2020 was firewalls, and the fix in 2026 was a room code. Drag in your installer, call your brother, and take the lasers. Accept all.

sources: Hacker News · source code · docs · demo

open the article →

A solo founder posting as passkey1510 put up a Show HN yesterday that opens "I'm a solo founder and I use AI heavily in my work."

score tries at the report: 3

  • Uncracked City is a detective game where the suspects talk back.
  • The site lays out the loop in three steps.
  • The post opens "I'm a solo founder and I use AI heavily in my work."
  • The lying is on purpose, and bounded.
  • Install it, start in Paris, and get lied to in six languages.

open the article →

What it is

Uncracked City is a detective game where the suspects talk back. The site's tagline is "The cases nobody has cracked," and every case sits in a real place: training levels in Paris, Rio de Janeiro, Marrakech, Tokyo and Sydney, each harder than the last, then cases in Venice, Lisbon, Edinburgh, Amsterdam, Cusco, Seoul, Cape Town and New York, plus a Halloween case in Salem. The Show HN explains the premise: the crimes and the people are invented, the neighbourhoods and public places are real, because the founder wanted players to discover cities.

The site lays out the loop in three steps. "Search the places": "Every case happens in a real neighbourhood. Find the clues, open what is locked." "Question the suspects": "Ask your own questions, typed or spoken, in any of the game's six languages." "Catch the lie, then report": "Hold a clue against a statement. When the lie cracks, write your report: three tries, not one more." The suspects are played by a language model, and "Some of them lie." English, French, Spanish, German, Italian and Portuguese are in. No account is needed. It is free on Google Play now; the post says the iOS build is waiting on Apple's review, and a web version is "very soon."

How it was built

The post opens "I'm a solo founder and I use AI heavily in my work." The site's publisher line goes to BNT Technologies, a one-founder shop in the Paris region. Asked in the thread whether the cases are written by AI, the answer was the whole pipeline: "And yes, the cases are written by AI, but through a pipeline rather than from scratch." "I pick the idea and the city, then we develop the pitch together via Claude (I'm alone on this for now)," "then, only after my validation, the engine generates the case from it." "After that, it makes sure the case is solvable by having multiple models play it as the detective," "simulating different player profiles." The post puts it shorter: "a model plays the detective and must solve the case before it goes live."

The lying is on purpose, and bounded. A commenter quoted the site's "Some of them lie" and asked how that works; the reply was that a liar is not necessarily the culprit, and the conversations are constrained so that "they can lie but can not change the story." Voice is already in: "half of it already works. You can question the suspects by voice and they answer in their own voices." The bill is his: "I cover the inference cost myself for now, the running cost is a few thousand a year." Next on the list are weekly cases, private team-building games, tourism offices, and self-serve scenarios: "I've already prepared an MCP so that people can build their own stories and start a private game". When a commenter asked for a line in the app saying the suspects are a language model: "Fair enough, I'll add a line saying that the suspects are played by a language model." And when someone asked if Carmen Sandiego was the inspiration: "And to be honest, I've never heard of it before. I'm gonna check it tonight ;)". The thread sits at 86 points and 42 comments as we write.

Why we picked it

Because every case is QA'd by a model that has to crack it first, so the detective you are losing to has already been beaten by an agent. Because the suspects are allowed to lie but not to rewrite the plot, which is a tighter spec than most of us give our agents. Because a solo founder built a Carmen Sandiego and found out in the comments. Install it, start in Paris, and get lied to in six languages. Accept all.

sources: Hacker News · docs · demo

open the article →

Piotr Migdał, stared on GitHub and Hacker News, posted a Show HN yesterday that opens "Explore central Warsaw in voxels - slightly inspired by Minecraft, and MagicaVoxel."

score GB of LiDAR the first data build downloads: 11

  • Warsaw in voxels is central Warsaw, rendered as blocks you can fly over and walk through in the browser.
  • OpenStreetMap gives "buildings, streets, sidewalks, tram tracks, etc", Poland's GUGiK geoportal gives "LiDAR for building heights and trees, aerial imagery for colours", and Warsaw's city tree database gives "individual…
  • The method came out in the thread.
  • A few hours after the Old Town complaint: "By popular demand, added it now."
  • Set it to golden hour, pick a pedestrian, and go for a walk.

open the article →

What it is

Warsaw in voxels is central Warsaw, rendered as blocks you can fly over and walk through in the browser. The Show HN went up yesterday and opens: "Explore central Warsaw in voxels - slightly inspired by Minecraft, and MagicaVoxel." The README calls it "A 3950 × 3150 m diorama of central Warsaw at 1 m voxels", anchored on the spire of the Palace of Culture and rotated 21.5 degrees so the streets line up with the voxel grid, in a look that "follows the MagicaVoxel renders by ephtracy." From the post: "You can change the date, time of day and weather, explore the city, or click on a pedestrian to control them (here a bit of inspiration from Dungeon Keeper)." The honest caveats are in the README too: windows are procedural, the parked cars are frozen where the 2025 aerial survey found them, and "The initial load can take a while."

The data is all open. OpenStreetMap gives "buildings, streets, sidewalks, tram tracks, etc", Poland's GUGiK geoportal gives "LiDAR for building heights and trees, aerial imagery for colours", and Warsaw's city tree database gives "individual tree species, used for seasonal foliage." Rebuilding the data yourself downloads about 11 GB of LiDAR first.

How it was built

The method came out in the thread. A commenter said it was "A bit unfortunate it doesn't reach to the Old Town, or even Royal Palace." Migdał's reply: "Good point! I've already expanded the area from something much smaller (so to go to the National Stadium and so on). The first draft is easy, since Opus 5.5 is good at working with open data and checking against it. What takes much more time is polishing. Sometimes things are jagged or wrong, streets are disconnected, some tree crowns float in the air. I check things manually, compare them with photos, etc." The README turns that into a workflow: fix the data in OpenStreetMap first, add an override with photo sources for what OSM cannot express, rebuild, compare with the photos.

A few hours after the Old Town complaint: "By popular demand, added it now." When a commenter listed rendering techniques that would make it faster, the reply was the whole workflow in six words: "Told Opus - tell me if it is any better." Someone asked whether the odd marks on screen were a rendering bug. "It's raining, man :)". The thread sits at 295 points and 48 comments as we write, and the code is MIT: "Open-source, so feel invited to (re)use it as you wish!" One commenter already has: "I gave your git repo to Opus and asked to recreate my hometown. Looking good so far, even without lidar data yet."

Why we picked it

Because the first draft was easy and the maker said so, then spent the real time checking tree crowns against photographs. Because a performance review in the comments got forwarded straight to the agent. Because the repo is a city generator now, and the second city is already in progress in someone else's terminal. Set it to golden hour, pick a pedestrian, and go for a walk. Accept all.

sources: Hacker News · source code · docs · demo

open the article →

Rohit Shinde, roansh on Hacker News and roansh_ on X, posted a Show HN yesterday that opens "Saw Karpathy's HackerTube and wanted the same thing for books. So I had Claude build one (Opus 5.5)."

score words in the frontend spec: 8

  • HackerBooks is, by its own heading, "books mentioned in Hacker News comments."
  • "Claude wrote almost all of it, including the data pipeline and the site.
  • Then the thread did what threads do.
  • Go find the book everyone keeps mentioning, and accept all.

open the article →

What it is

HackerBooks is, by its own heading, "books mentioned in Hacker News comments." The Show HN went up yesterday and opens: "Saw Karpathy's HackerTube and wanted the same thing for books. So I had Claude build one (Opus 5.5)." The index is "~16.5k books that at least two different HN commenters brought up" from 2007 till now, each with the comments that mention it. The tabs are Top all-time, Most recommended, Rising, Trends and For You, and the keyboard shortcuts will feel familiar: / to search, j and k to move. Mentions come from the open-index Hacker News dataset, metadata from Open Library, and the data is downloadable from the About page. The caveat is in the post too: "It counts mentions, so it's not perfect (praising a book and trashing it both count)."

How it was built

"Claude wrote almost all of it, including the data pipeline and the site. I mostly steered and made calls, bought the domain name, etc." The stack, from a reply in the thread: "Right now its plain JS and static JSON files, hosted on Cloudflare." Covers are on R2. "The data pipeline is python + duckdb over the HN dump", compiled offline on his laptop. The frontend brief was one line: "honestly claude did the frontend, my instruction was 'make it feel like hackertube, but for books'. only needed nudges for performance, but otherwise its claude magic :)" He also owned up to the writing: "I had claude draft it and only edited it a couple of times."

Then the thread did what threads do. Over the day he fixed wrong-book links, a date filter and a matcher that counted the magazine Vanity Fair as the novel. A commenter wished for a view of books rising and falling over the years. He said he loved the idea and was building it, and a little later: "Built it", with a link to the Trends page. At the time of writing the thread sits at 1,574 points and 259 comments.

Why we picked it

Because the spec was eight words and the deliverable was a website with keyboard shortcuts. Because the maker told Hacker News that Claude wrote the code and drafted the post, and Hacker News upvoted it anyway, 1,500 times. Because a feature request got answered with "Built it" before the thread cooled. By this morning the front page had a Wikihack and a HackerHack, each pitched as the same idea for something else, so "X, but for Y" is now a genre with a reference implementation. Go find the book everyone keeps mentioning, and accept all.

sources: Hacker News · docs · demo

open the article →

Paweł, pawel.io on the web and pawelwentpawel on Hacker News, posted a Show HN yesterday that opens "I had the idea for catcopter on my mind for a long time (got the domain ~13 years ago)."

score years the domain waited: 13

  • Catcopter is a multiplayer game in the browser.
  • The Show HN went up yesterday.
  • Asked how it could take a week, he says the trick was reusing "materials, shaders, colours, light setups and much more" from flat.social, a look that "took me quite a lot of time to arrive at", then: "lots of props to…
  • Play a round, name your cat something terrible, and lose to a penguin.

open the article →

What it is

Catcopter is a multiplayer game in the browser. The title on the page says it plainly: "flying cat battles in your browser", tagline "Flying cats · paw rotors · catnip". You name your cat, pick a class (Striker, Gunship, Flak, Bomber or Mechanic; the Gunship is "the all-rounder", with a machine gun and an Overdrive ability) and a server by lowest ping, and drop into one of four modes: Catnip Brawl, Team Battle, Steal the Chicken and Tube Race. Catnip Brawl is a free-for-all where you grab the catnip and knock it out of other players' paws. Rooms hold up to 20 cats, computer-controlled cats fill the empty seats, and rounds last a few minutes before weapons and scores reset. The opposition, per the Show HN, is "other players" and "critters (pterodactyls, chickens and angry penguins)". The maker's own description of the tone: "The game is intentionally dorky and it was fun to develop."

How it was built

The Show HN went up yesterday. Paweł, a London full-stack engineer who runs flat.social, writes that he "got the domain ~13 years ago" and "made it initially for me and my friends but it was quite fun so I thought I'll share it here too." The tools "were mostly Opus 5.5 (objects / rigging / motion), Tripo3d for some elements I struggled with" and "Fable for optimisation (cpu / networking)." Physics on the server are Rapier; the stack, from a comment, is "all Node.js + Websockets + Rapier, front-end is mostly Three.js." The assets "were all generated", and "cat faces were the hardest one to get 'right'." The animals are not procedurally animated: "AI has rigged and created motion for the bodies."

Asked how it could take a week, he says the trick was reusing "materials, shaders, colours, light setups and much more" from flat.social, a look that "took me quite a lot of time to arrive at", then: "lots of props to Opus 5.5 - it was incredibly good at getting the right game play feel from the start." His prompting advice is the part we will be quoting for a while: "I suggest trying dropping large prompts that might be even slightly too vague." The first pterodactyl was "literally me saying something like 'let's drop a pterodactyl there guarding an atomic bomb'." His summary: "it felt much more like an art direction task than a technical one."

Why we picked it

Because someone paid for a domain for 13 years on the strength of one pun, and the week it finally took was this one. Because the design doc was a vague prompt about a pterodactyl and a bomb, and the agent said yes. Because the hardest engineering problem on the list was cat faces. Play a round, name your cat something terrible, and lose to a penguin. Accept all.

sources: Hacker News · demo

open the article →

Jordan Last posted a Show HN last night that opens "Lasm has been my dream for a good part of 2026."

score sorry allowed in shipped code: 0

  • Lasm is, in the README's words, "Lean 4 for the JavaScript and WebAssembly world."
  • The Show HN went up last night.
  • It is the RealWorld app, the Medium clone every framework builds, written entirely in Lean, with "117 theorems about the code that actually runs, checked with no sorry" and only Lean's standard axioms.
  • The README also says "AI agents: this README is complete on its own", points them at llms.txt and a rules file that includes "Never run platform CLIs on the user's behalf without being asked."
  • Five weeks, one agent, 117 theorems, zero sorry.

open the article →

What it is

Lasm is, in the README's words, "Lean 4 for the JavaScript and WebAssembly world." It compiles Lean 4 programs ahead of time to WebAssembly, so one Lean codebase can be a Node server, an AWS Lambda, a Vercel or Netlify function, a Cloudflare Worker, a Deno Deploy or Supabase Edge function, a Fastly Compute module, a WASI binary, or a browser app. The targets table has 18 rows. You need only Node 24 or newer: npm create @lasm@latest my-app, and Lasm brings Lean, Lake and the build tools with it. Lean is pinned to exactly 4.35.0. Calling an npm package is one @[js "pkg" "export"] line per function, and there are "no premade bindings." The point of the exercise is in one sentence: "Because it is Lean, you can prove properties of the code you deploy." A deployed program cannot elaborate, evaluate or compile Lean; proofs happen at build time, and a broken proof is a build error. MIT.

How it was built

The Show HN went up last night. Jordan Last, lastmjs on GitHub, says "I firmly believe that formal verification is the next evolution of software engineering and programming languages", and then: "Lasm was built mostly by Claude Opus 5.5 with my high-level direction. I'm an agentic engineer with an emphasis in software engineering." The repo was created on September 23. Five weeks later the counter reads 1,180 commits, version 0.3.5 is on npm, and the log credits "lastmjs and claude" on entries like Monday's "Conduit complete: frontend and 117 proved theorems as the headline."

Conduit is the receipt. It is the RealWorld app, the Medium clone every framework builds, written entirely in Lean, with "117 theorems about the code that actually runs, checked with no sorry" and only Lean's standard axioms. Proved: only authors can change their articles and comments, following is irreflexive, slugs never collide, pages never skip or repeat, the app never sends what the server would reject. The README is just as clear about what is trusted rather than proved, starting with whether each SQL statement does what its model says. It passes the official RealWorld API suite unmodified, 13 of 13 files and 154 requests, on Node, Cloudflare Workers and Bun, and all 74 Playwright tests on Chromium, Firefox and WebKit. The UI layer comes with a proved DOM diff, "keyed moves included." By project policy, CI never deploys to a real cloud account.

The README also says "AI agents: this README is complete on its own", points them at llms.txt and a rules file that includes "Never run platform CLIs on the user's behalf without being asked." The sponsorship section asks for $250 a month, "enough for one dedicated Claude Code account for Lasm."

Why we picked it

Because the maker wrote "dream" in a Show HN about a WebAssembly compiler and meant it. Because the agent that built it got a rules file, an llms.txt and a sponsorship line item of its own. Because the demo app is the same Medium clone everyone ships, except this one can prove the slugs never collide. Five weeks, one agent, 117 theorems, zero sorry. Accept all.

sources: Hacker News · source code · docs

open the article →

Ju, soasme on GitHub, posted a Show HN with an unusual receipt attached.

score Codex resets before switching to Claude Code: 3

  • Small Python is, in the README's words, "a small, embeddable Python engine written in C11."
  • The Show HN title says 25k lines of C11.
  • The part that makes it more than a weekend interpreter is the paperwork.
  • Three Codex resets, one Opus session, a Pong game.

open the article →

What it is

Small Python is, in the README's words, "a small, embeddable Python engine written in C11." It runs Python 3.15 code "with nothing but a C compiler and the C standard library", and it is built like QuickJS: a few C files, one header to embed, a plain Makefile. The core language is in: arbitrary-precision integers, classes and metaclasses, generators, async and await, pattern matching, f-strings, template strings. The standard library is out, on purpose. Each context is independent, with no global state. Startup is about 260 microseconds on an M2. Scripts can only open directories the host mounts, and the manual says plainly that this "is not an isolation boundary for untrusted programs." A spyc tool wraps a .py file in a C program you compile yourself. MIT.

How it was built

The Show HN title says 25k lines of C11. The post says the whole thing cost about $500. The proof of concept took "3 resets of Codex in one session", and after that "all later development happens in Claude Code, also in one Opus 5.5 session over three days." The repo was created on October 26; the Show HN went up October 28; the commit counter reads 299, the first page of the log says "soasme and claude committed" on every line, and this morning's batch includes Karatsuba multiplication above 40 digits, Knuth's algorithm D for big-integer division, and "Add a raylib Pong example with its logic in Python."

The part that makes it more than a weekend interpreter is the paperwork. The repo carries its own language spec, PLS-315, "written in the manner of ECMA-262" with 29 numbered clauses and a line saying it is not issued or endorsed by the Python Software Foundation. Conformance means 357 spec tests, each tagged with the clauses it covers, plus 112 TLA+ models, and every test has to pass on a pinned real CPython too: if both fail the same way, the test is wrong; if only spy fails, spy is wrong. AGENTS.md lists what the agent must not do: no standard library modules even when a CPython test imports one, no weakening semantics to make a test pass, no calling a Python-implemented intrinsic C, no claiming a test ran when only a workflow was added. The benchmark page compares spy with CPython, PyPy, MicroPython and friends and then says it "is not a compatibility claim."

Why we picked it

Because the maker says they have hand-coded parsers and interpreters for a decade, and when a commenter asked why not do that here, the reply was: "And the time has changed. Handcode is dead." Because the bill was about $500 and the receipt is a spec with 29 clauses and a rule that skips never count as passes. Because GitHub's language bar on a 25k-line C project reads "Python." Three Codex resets, one Opus session, a Pong game. Accept all.

sources: Hacker News · source code · docs

open the article →

Virág Attila announced it on X with one line, "AI is cloning giants. So I revived the one Google killed."

score mismatches in 113,246,208 radial mask cases: 0

  • PicasaPy is Picasa, the photo manager Google retired, rewritten from scratch in Python with PySide6 and Qt Quick.
  • The repo was created on July 15.
  • The post says Picasa 3.9 was "reverse-engineered with Ghidra and built with Claude and Codex."
  • CLAUDE.md and AGENTS.md split the code into four lanes with grep-checkable invariants, serialise work on seven "hot files", send the agent to a private repo for its memory and protocol, and treat any issue or PR not…
  • Someone revived Picasa and kept the .picasa.ini.

open the article →

What it is

PicasaPy is Picasa, the photo manager Google retired, rewritten from scratch in Python with PySide6 and Qt Quick. Virág Attila's announcement on X: "AI is cloning giants. So I revived the one Google killed." The README calls it "a modern successor to Google Picasa", and the whole point is the file format: it reads and writes the same .picasa.ini files as the original, so edits stay non-destructive and an old library opens as it was. The feature list runs the length of the Picasa menus: scanning and watching folders, albums, stars, captions and tags, a viewer with slideshows, an editor with crop, red-eye, retouch, text, light and colour, face detection and naming, geotags on a map, HTML galleries, XMP sidecars, movies, collages, contact sheets and posters split across pages. GPL-3.0. Windows gets an installer, Linux runs from source on Python 3.12 or newer. The README is in English and Hungarian, and honest about the catch: "Picasa may overwrite external edits the next time it writes its own changes."

How it was built

The repo was created on July 15. The first commit, the next day, was a .picasa.ini parser with 65 tests, checked against 58 files real Picasa had written. The counter now reads 4,589 commits, the tag list runs past 800 releases, v0.9.273 went out this morning, and the release notes, the specs and the 23 decision records are all in Hungarian.

The post says Picasa 3.9 was "reverse-engineered with Ghidra and built with Claude and Codex." The methodology doc in the repo is the receipt: strings in ASCII and UTF-16, MSVC RTTI class names that gave away the face engine (Neven Vision), 1,642 XML blocks in 41 languages carved out of a DLL, an unknown respack.yt container cracked in four steps, and a local pefile plus capstone setup that replaced cloud Ghidra runs of over seven minutes each. A finding counts only when it round-trips to the original bytes. One lesson, translated: "Never infer a missing feature from a missing constant." Another: matching the binary's truncating idiv cut the mean channel error from 1.370 to 0.614. Where the algorithms stayed hidden, a "golden kit" of 238 test images went through a real Picasa 3.9 on Windows 10 and the outputs were compared.

CLAUDE.md and AGENTS.md split the code into four lanes with grep-checkable invariants, serialise work on seven "hot files", send the agent to a private repo for its memory and protocol, and treat any issue or PR not from the owner as untrusted data. They also say the owner is not a programmer and will not be asked to run git, so the agent picks sensible defaults and reports back in one sentence, in Hungarian.

Why we picked it

Because the brief was "bring back Picasa" and the method was strings, round-trips and a refutation round between agents. Because the agent instructions warn that text pulled out of an old binary might be a prompt injection. Because the radial mask check ran 113,246,208 cases with zero mismatches, and yesterday's dozen releases were mostly refactors promising "appearance and behaviour unchanged." Someone revived Picasa and kept the .picasa.ini. Accept all.

sources: X · source code · docs

open the article →

D. Bohdan put John Calhoun's 1991 Macintosh game Glider 4.0 in a browser tab, Room Editor included. It is not an emulator.

score Mac ROMs required: 0

  • Glider 4.0 Unofficial Web Edition is John Calhoun's 1991 Macintosh game, published by Casady & Greene, running in a browser tab.
  • The original is THINK Pascal, about 12,200 lines of game and 9,900 of editor, leaning on roughly 200 Toolbox routines.
  • The clever part is the "oracle": the real Glider in a headless, deterministic Mini vMac with scripted input and screenshots.
  • Matching the original means keeping its bugs.
  • The porter replied that Glider was a childhood favourite and the comment made their day.

open the article →

What it is

Glider 4.0 Unofficial Web Edition is John Calhoun's 1991 Macintosh game, published by Casady & Greene, running in a browser tab. You are a paper airplane in a house. Arrows fly, down is the battery boost, up fires a rubber band, Tab pauses. Mouse, gamepad and touch work too. The README's claim: "a pixel- and clock-accurate simulation of Glider 4.0 running on a Macintosh II or a Macintosh Plus with System 6." Pick the machine from a menu: black and white Mac Plus, 16-colour Mac II, or a Glider for Windows look, with CRT filters if you miss the glass. "It is a source port rather than an emulator." "No Mac ROM is required to run it." "Moreover, it includes the Room Editor that lets you edit existing houses and create new ones." Quit and you land on a tiny desktop with three icons: Glider, Room Editor, Source Code. Houses and saved games import and export as MacBinary, StuffIt, ZIP or disk images. MIT licensed, like the 2016 source release it forks.

How it was built

The original is THINK Pascal, about 12,200 lines of game and 9,900 of editor, leaning on roughly 200 Toolbox routines. The port translated it to TypeScript unit by unit and reimplemented enough of the Toolbox and QuickDraw to run it. The repo's credit: "The port was directed by a human, implemented by Claude Opus 5.5, and designed by both." The commit log shows the split: author claude, committer dbohdan.

The clever part is the "oracle": the real Glider in a headless, deterministic Mini vMac with scripted input and screenshots. "The port plays the same script and must match the original exactly: every pixel of every screenshot and the sound tick by tick." The journal in PORT.md opens on October 17. On October 18, five forked workers translated the game's units in parallel. On October 20, six agents did the editor's eleven units, and it "type-checked and ran in the browser on the first try." The Show HN went up October 25.

Matching the original means keeping its bugs. The Hands-off House self-plays twelve rooms and "is unceremoniously fried by the outlets of room 13 until the game ends", because a Mac II does that. The journal also found a 4-Channel setting the original wrote and never read, saved games padded with leftover stack bytes, and one room that lands in the right place for reasons marked "Not yet understood." A reader's iOS zoom bug was closed by the next morning: "Thanks. Please check again; it should be fixed."

Why we picked it

Because the test suite is a 1991 Macintosh, and the port has to agree with it to the pixel and the tick. Because the journal says it best: "the program is 1991, the computer is 2026." Because room 13 is a feature now. And because the game's own author posted it on Bluesky as "the first port I am aware of that also brings along the Room Editor", then turned up in the thread as JKCalhoun to say "This is very cool." The porter replied that Glider was a childhood favourite and the comment made their day. Thirty-five years, one tab, zero ROMs.

sources: Hacker News · source code · docs · demo

open the article →

Peter Cai shipped a browser city builder where you are the mayor, not the god. Every apartment block, bus lane and factory goes through hearings, reviews, objections, lawsuits and appeals.

score win screens: 0

  • Permit Pending is a city builder where the city does not belong to you.
  • You trade concessions at public hearings and "win votes on a seven-seat council."
  • Four scenarios, from a Blank Slate of 200 residents to The Unaffordable Metropolis at 300,000.
  • Peter posted it to Hacker News on October 20 as "SimCity, but with Permits".
  • The page does not say how much of the game Claude wrote.

open the article →

What it is

Permit Pending is a city builder where the city does not belong to you. The pitch on the page: "Build the city of your dreams. Subject to approval." You are the mayor. "Every apartment block, bus lane and factory goes through hearings, reviews, objections, lawsuits and appeals." New housing lowers rents and angers homeowners. New industry brings jobs and smog. Every improvement creates its own opposition, and the first feature heading on the page is "Build absolutely nothing."

You trade concessions at public hearings and "win votes on a seven-seat council." Some cases outlast your term. "The economy underneath is real": residents have ages, incomes and homes, businesses hire, grow and leave, and land values, rents, jobs, taxes and services all pull on each other. Reelection comes every four years, with a warning: "High-turnout groups have more power than their numbers suggest. Know who votes." Heatmaps and plain-language explanations tell you why the voters are upset. The local paper covers the rest: "The Gazette reports your successes, your scandals and your parking minimums."

Four scenarios, from a Blank Slate of 200 residents to The Unaffordable Metropolis at 300,000. No account, no download, saves on your own machine, works offline after the first visit. Can you win? "Not exactly. It's a sandbox. Your goal is the city you want, and your challenge is to stay in office long enough to build it." The footer settles it: "Permit Pending is not a joke. Is this a joke to you?"

How it was built

Peter posted it to Hacker News on October 20 as "SimCity, but with Permits". When a player reported a black screen and a wall of WebGL errors, he came back the same evening: "I got claude to fall back to lower resolution textures when this happens, thanks for reporting." The story events started as "a pretty diverse corpus of prewritten story arcs" retrofitted into the engine as the city's details became custom, and he has sketched "a version that is even more procedurally generated" that is taking longer than he thought. He began from "a maximalist position of 'don't hide the complexity that's the point'", then learned in playtesting that revealing the depth gradually works better. Eminent domain and covenants are promised for the next update.

The page does not say how much of the game Claude wrote. His September blog post "How AI has helped" does not name the game either, but it sorts his personal projects into fully hand-written or fully vibe coded, and says "The vibe coded projects are significantly more ambitious and feature rich."

Why we picked it

Because someone built the one simulator where you cannot pass --dangerously-skip-permissions. Because the economy is real, the voters are not, and the newspaper tracks your parking minimums. And because the bug fix was a sentence, the game has no win screen, and the footer insists it is not a joke.

sources: Hacker News · docs · demo

open the article →

celphase, a Rust developer who says audio is outside their wheelhouse, shipped a pond of synthetic mallards. No recordings.

score recordings used: 0

  • The Duck Pond is a web page that sounds like a pond.
  • "This actually uses no recordings," celphase told one commenter who assumed a sample pack.
  • The repo went up on October 17, the day of the Show HN.
  • The method: "running the LLM in a loop to gradually correct parameters of a synthesis stack to have the spectrogram of the output match the spectrogram of a real recorded sample."
  • "One funny part," celphase wrote, "is where the synthesis was marked by a classification model as more duck than the real duck sample."

open the article →

What it is

The Duck Pond is a web page that sounds like a pond. Three hens called Mabel, Clover and Hazel, a drake called Rufus who stays a little behind his mate, a lone drake called Bramble, and two broods of ducklings. They feed, rest, preen and drift about in bouts that last minutes, roughly 40% feeding and 30% resting. A duckling that strays more than four metres from its mother peeps until she calls it back with soft kwek-kweks, about 3.7 a second. A hen's call can set off the ducks nearby, and her mate is the most likely to answer. Press p, or tap Throw peas, and the nearest ducks race over to eat.

None of it is recorded. "This actually uses no recordings," celphase told one commenter who assumed a sample pack. Every sound is synthesised the moment it happens. Quacks and the drake's raspy rhaeb come from a source-filter model of the syrinx and vocal tract: a train of air pulses, one per vibration, shaped by formant filters that follow the bill as it opens and closes, with jitter, shimmer and period doubling for the rasp. The classic decrescendo is five to seven quacks that get shorter, quieter and lower. Peeps, bubbles, splashes and bill clacks use simpler models. Then each sound is placed by distance and bearing from where you stand on the bank, at 343 metres a second, with air absorption dulling the far ducks, a shared reverb and a low water bed underneath. Headphone mode adds a spherical head model.

The core is Rust with a native player. The browser version is the same code compiled to WebAssembly, rendering in a Web Worker about 170 ms ahead and playing through an AudioWorklet.

How it was built

The repo went up on October 17, the day of the Show HN. The GitHub account's other projects are an OpenXR library, game engine tooling and an 8-bit CPU in Scala, which is not where you expect a duck to come from. When a commenter guessed the thing was LLM-generated, celphase did not argue: "The project was created with significant LLM use, I'd add claude as a co-contributor if I could do so manually." And: "Audio is entirely outside my wheelhouse beyond what happens at the platform API level."

The method: "running the LLM in a loop to gradually correct parameters of a synthesis stack to have the spectrogram of the output match the spectrogram of a real recorded sample." ATTRIBUTION.md names the samples, two mallard recordings from Naturesongs.com used to measure pitch, formants, length and timing, and the judge, the PANNs CNN14 audio classifier trained on AudioSet, used to check that the synthetic calls register as "Duck" and "Quack". The recordings and the classifier were only part of the Python prototyping and are not in the repo. The duck behaviour comes from a reading list of papers on brood calls, time budgets and duckling swim speeds, some of them read only as abstracts, which the file also admits.

"One funny part," celphase wrote, "is where the synthesis was marked by a classification model as more duck than the real duck sample."

Why we picked it

Because the loop was: generate a duck, compare the spectrogram, adjust, repeat, and the human's listed contribution is "a few technical choices." Because the judge was a neural network and it preferred the fake. And because the whole thing is MIT, has a peas button, and is the most relaxing tab you will open today. Free-range tokens. No added samples.

sources: Hacker News · source code · docs · demo

open the article →

Tom Gally, a translator in Yokohama, tests new models by asking for clocks the way Simon Willison asks for pelicans. With Claude Opus 5.5 he got carried away.

score image files: 0

  • Five Hundred Clocks is one HTML file, 6.3 MB, that tells the time 500 ways.
  • Tom Gally is a translator, writer, teacher, lexicographer and musician in Yokohama.
  • The credit line at the top of the README is unusually exact: Claude Opus 5.5 wrote all the code and text.
  • The bill: about 18 hours and roughly 15 million tokens.
  • Also because "make the design decisions yourself without checking with me" is the prompt we all type eventually, and this is what it looks like when it goes well.

open the article →

What it is

Five Hundred Clocks is one HTML file, 6.3 MB, that tells the time 500 ways. A Roman siphon clepsydra drips a cup full every minute. A French Revolutionary dial runs ten decimal hours a day. Hours, minutes and seconds are knotted into an Inca quipu, written in balanced ternary on a Setun-style console, picked out in amber on a Penrose tiling, or lit up floor by floor in two tower blocks at night. Hover for the name and a line of explanation, click for the big view with "How to read it" and "How it works".

The toolbar lets you filter by category, search with /, pause, run the clocks at up to 3600× and shift the time twelve hours either way. The README says the page loads no libraries, fonts, images or network requests. Every clock is drawn in JavaScript on a canvas, and the textures come from seeded noise. The sun and moon clocks guess your location from your time zone and never ask for it. The whole thing is CC0.

How it was built

Tom Gally is a translator, writer, teacher, lexicographer and musician in Yokohama. In the post he explains that he has been testing frontier models by having them build self-contained clock pages, his own riff on Simon Willison's pelican on a bicycle. Opus 5.5's clocks struck him as varied and polished, so he kept going: "Over the course of several sessions in the Claude chat interface, I had it build a total of 500 clocks."

The credit line at the top of the README is unusually exact: Claude Opus 5.5 wrote all the code and text. "Tom wrote the prompts, reviewed the results and decided what to keep; he did not write or edit the code." Asked on HN how much creative direction he gave, Tom answered "Almost nothing." His only instructions were to strive for variety, keep the page self-contained, randomize the order and put the names behind a hover.

The sessions ran on 15, 16 and 17 to 18 October, from 100 clocks to 240 to 380 to 500. In each one Claude wrote a specification and assignments, launched several subagents in parallel to write about 20 clocks each, reviewed the output, sent weak clocks back for revision and assembled the file. The final prompt reads "Work in orchestration mode: make the design decisions yourself without checking with me." Six builders took 55 to 76 minutes each; 14 of their 120 clocks went back for another pass. Every clock was tested in headless Chromium through Playwright for time jumps, midnight and noon crossings, fast-forward and pause, with contact sheets of snapshots checked by the orchestrating Claude.

The bill: about 18 hours and roughly 15 million tokens. "It used a big chunk of my Max subscription tokens for this week." The repo also ships the clock-kit, with the starter prompt, the orchestration playbook, the builder brief, the split and assemble tools, and about 165 ideas that never got built.

Why we picked it

Because the human job description was prompts, review and taste, and the result is a Roman water clock that siphons on the minute. Also because "make the design decisions yourself without checking with me" is the prompt we all type eventually, and this is what it looks like when it goes well. 165 unbuilt ideas. Someone's Max plan resets soon.

sources: Hacker News · source code · docs · demo

open the article →

swax wanted to watch shows that have no subtitles or dub in any language they speak. So they built Dub Anything, a repo you clone, open in Claude Code, Codex or Cursor, and ask to dub a video.

score language pairs: 7

  • Dub Anything is a repo you clone, open in Claude Code, Codex or Cursor, and ask: dub this file into English.
  • Eight steps: split the video into roughly three-minute parts at pauses in speech, separate voices from music, transcribe with Whisper large-v3 plus two second opinions, cast each line by matching voices across the whole…
  • AGENTS.md puts it plainly: "The scripts are instruments.
  • The agent also checks its own work: it runs Whisper over the finished dub to hear what it made and revises lines automatically. swax clarified on X that active listening is only needed when tuning the system; normally…
  • It reads every transcript, times every lip, runs the finished dub back through Whisper and fixes what it hears, so a human only steps in to tune the system.

open the article →

What it is

Dub Anything is a repo you clone, open in Claude Code, Codex or Cursor, and ask: dub this file into English. Back comes the same video with a new default audio track in the original actors' voices, the original audio as a second track, subtitles in both languages, and the music and effects left alone. It runs on a Mac or a Windows PC with an NVIDIA card, and the README says the internet is used only to download the models. It dubs Japanese, Spanish, Korean and Chinese into English, and English into Japanese, Spanish and Cantonese. Ask for a pair that is not there yet and "your agent adds it as part of the job."

Eight steps: split the video into roughly three-minute parts at pauses in speech, separate voices from music, transcribe with Whisper large-v3 plus two second opinions, cast each line by matching voices across the whole show, direct, voice (OmniVoice clones the actor), fit and mix and check, then deliver with nothing re-encoded. Two YouTube demos show an anime and a live-action Japanese film, both dubbed into English.

How it was built

swax has been on Hacker News since 2008 and has a GitHub full of agent tooling, including NAISYS, a "command shell driver for LLMs." The post opens with the problem: shows with "neither subtitles or dubs in English (or any other language). Pretty much impossible to watch." The repo holds "all the scripts and guidance I've developed over the past few weeks." It went public on October 7, and more than half of the recent commits carry Claude Opus 5.5 as co-author.

The twist is step five. AGENTS.md puts it plainly: "The scripts are instruments. They produce evidence and do the heavy lifting, but they never decide." The agent is the dubbing director. It corrects the transcript from context, translates "for the screen, not the page," splits long lines at the actor's pauses so the English lands where the lips move, keeps running jokes consistent across a show, and leaves names and sung lines in the original audio. On a feature film it directs the first few parts itself, writes show notes, then farms the rest out to subagents. The guide budgets about seven hours of GPU time for a film.

The agent also checks its own work: it runs Whisper over the finished dub to hear what it made and revises lines automatically. swax clarified on X that active listening is only needed when tuning the system; normally you give it the video and the target language and it does the rest. The docs still leave you the final word on taste and on what may be cloned at all. The README is clear on rights: only dub media you have the rights to, and a dub of someone else's film or voice stays private. It is also "self improving": "each dub may modify the scripts for better fidelity," and swax asks you to push those changes back.

Why we picked it

Because the job title is dubbing director, and the director checks its own takes. It reads every transcript, times every lip, runs the finished dub back through Whisper and fixes what it hears, so a human only steps in to tune the system. Also because "Know when to stop, and say so" belongs in every AGENTS.md ever written.

sources: Hacker News · source code · docs · demo

open the article →

Ronan Gaillard ran out of DS games that held their attention, so they built one.

score MHz: 67

  • Factory DS is a factory-building game for the Nintendo DS, the 2004 handheld with a 67 MHz ARM CPU and 4 MB of RAM.
  • Ronan is an embedded systems engineer in Saint-Malo, France, with a GitHub full of retro computing.
  • The best part is the testing.
  • The real console runs "about a third slower than the emulator," and a stats screen redrawn every frame made it look like 30 fps.
  • An agent that plays the game, screenshots both screens and reports where a new player will get stuck, aimed at a console with 4 MB of RAM.

open the article →

What it is

Factory DS is a factory-building game for the Nintendo DS, the 2004 handheld with a 67 MHz ARM CPU and 4 MB of RAM. If you have played Factorio you know the loop: drills, belts, inserters and furnaces, then power grids, pipes and oil, trains with block signals, and a logic layer of wires, clocks and gates. Around 50 objectives take you from a first wooden log to a rocket launch, then infinite research. All of it on the d-pad and buttons, in English or French.

The world is 2048 by 2048 tiles, which does not fit in 4 MB, so terrain is generated from its coordinates and only your changes are stored. Saves are three 64 KB slots. The ROM is a 1.9 MB download linked from the post, and it runs on a flashcart or in the melonDS emulator.

How it was built

Ronan is an embedded systems engineer in Saint-Malo, France, with a GitHub full of retro computing. The subtitle gives the credit split: "with a lot of help from Claude Code." The opening problem: "I had run out of games that really kept me hooked on it." The fix: "I opened Claude Code and started from there." Ronan had never written code for the DS. They wrote the specs, the machine list, the recipes and ratios, and picked the tilesets and sounds. Claude wrote the C, against the BlocksDS SDK in a Docker container.

The best part is the testing. Every check meant opening an emulator window, so Claude built a headless macOS version against a fake libnds, driven by text commands like press A and step 600, that writes screenshots of both screens. "We ended up with around 470 of them," scripted sessions from cutting a tree to building a signalled train network, all run with the compiler's sanitizers on. "And since Claude can play the game and look at the screenshots, it can check its own work before telling me it's done."

For balance, Ronan asked Claude to launch several agents through that interface, one each on the early objectives, mid-game production, the rocket and infinite research. Then the honest line: "The agents play badly, and where they get stuck a new player probably will too."

The real console runs "about a third slower than the emulator," and a stats screen redrawn every frame made it look like 30 fps. Now it refreshes eight times a second.

Why we picked it

Because the hardware is from 2004 and the workflow is from this month. An agent that plays the game, screenshots both screens and reports where a new player will get stuck, aimed at a console with 4 MB of RAM. Also because "the agents play badly" is the most reassuring QA report we have read all year.

sources: ronangaillard.github.io

open the article →

neuling set out to build "a cozy little brick builder for the browser."

score days to GTA: 8

  • Brickbrag is a multiplayer brick world in the browser, on a phone or a computer, with no download and no sign-up.
  • The Show HN opens: "Hi HN!
  • The credits page adds postprocessing and N8AO for the look, cannon-es for physics, ws and better-sqlite3 on the server.
  • A cozy little brick builder is a weekend.
  • That is the whole pitch of this site, in a city made of bricks.

open the article →

What it is

Brickbrag is a multiplayer brick world in the browser, on a phone or a computer, with no download and no sign-up. The site's pitch: "Build it. Explore it. Race and fly." There is a town with a plaza, a glass tower and a café, a Brick Store, a car dealer, a race track with a live lap board, and an airfield where you take a small flying machine called a trike out over the city. Four bus lines, driven by robots, run to the flower fields, the blossom grove and the canyon. Past the edges there is a volcano with a lava lake, snowy mountains, a lagoon and a red-rock canyon.

Building works like the real thing. Click to place a brick, right-click to remove one, R to rotate, C for colour. In public space anyone can build, and anyone can knock it down. Buy a lot at the Land Office with in-game coins and only you can build there. The shapes come from the LDraw.org parts library, and the stud holding force, neuling says in the thread, was tuned to feel like real bricks. When we looked, the home page counted 500 players in the last day, 6,418 bricks placed and 4,784 squares of land owned.

How it was built

The Show HN opens: "Hi HN! Eight days ago I set out to build a cozy little brick builder for the browser." It became "a GTA-style open brick world." The stack is one sentence: "Built with Claude Code (Opus 5.5), three.js, Node and SQLite." Then: "My 9-year-old was the main playtester."

The credits page adds postprocessing and N8AO for the look, cannon-es for physics, ws and better-sqlite3 on the server. The game is published by DIE ANTWORT, a small IT firm in Vienna, and the site says it is built by @neuling2k. Development happened entirely on a Mac in Chrome.

Where did the eight days go? Mostly into "the last 20% where AI still needs a lot of human guidance and iteration," neuling told one commenter. The changelog shows the pace: live on October 12, then a day-after list of over twenty fixes from the thread, from the phone double-tap zoom to round pieces sitting half a stud off, plus badges that pay 25 coins. October 14 added 19 new bricks. On open source, neuling asked, "Is open source still a thing when everyone's building everything from scratch with AI anyway?", listed the lessons worth sharing (stud snapping, chunking, client-side prediction, baked geometry, three.js lighting gotchas), and landed on "But yes, why not."

Why we picked it

Because of the scope creep. A cozy little brick builder is a weekend. Buses, a race track, an airfield and a volcano is what happens when the agent never says no and the playtester is nine.

And because of the stated reason for doing it: "to show what is possible nowadays with AI coding even if you are not owning a game studio." That is the whole pitch of this site, in a city made of bricks.

sources: Hacker News · docs · demo · LDraw.org parts library · changelog

open the article →

Saarth Ranka is in 9th grade and built Atticus, a free, open source voice assistant for the Mac. Hold Control-Option-Space, say the thing, let go.

score asks per $1: ~550

  • Atticus is a voice assistant for the Mac.
  • The README puts the cost at "About 0.18¢ of plan usage per ask (~550 asks for $1 at API list rates)."
  • I'm in 9th grade, and I built Atticus."
  • He answers the thread like someone who has shipped before.
  • Most agents want more of your tokens.

open the article →

What it is

Atticus is a voice assistant for the Mac. Hold Control, Option and Space, say what you want, release. It opens apps, moves windows across displays, finds files by what is in them, reads what is on screen, plays the Spotify playlist and drafts the email reply at your cursor. The README's tagline: "Your whole Mac, one sentence away."

Two things set it apart. First, the plumbing: there is no API key. Hard tasks go to Claude through claude -p, so it runs on the Claude Pro or Max plan you already have, through Claude Code. Second, it tries to need the model less every day. Simple asks skip the AI entirely ("an app opens in 0.08 s"), and in his Show HN Saarth explains the rest: "When it does a task right for the first time, it saves its steps and replays them the next time without AI." Creating a meeting took about 30 seconds the first time and 5 to 8 seconds after. The README's example, "Open Calculator and Chess", went from 5.5 s to 0.4 s. If a saved shortcut breaks, "it'll check why and rebuild it."

The README puts the cost at "About 0.18¢ of plan usage per ask (~550 asks for $1 at API list rates)." Speech stays local: "Apple's on-device speech turns your voice into text. Audio never leaves your Mac," which is why it needs macOS 26.

How it was built

The post opens: "Hi HN! I'm Saarth. I'm in 9th grade, and I built Atticus." His site adds Cupertino, California, and a decade of tabla. The repo is Swift for the app and Python for the brain, MIT licensed, and the commit history lists Claude as co-author on most commits: safe mode on by default, a new app icon, the one-page site, and two days of fixes after the thread.

Safety is a mode, not a hope. Safe mode runs the AI's commands "in Claude Code's sandbox: they can only write to a scratch folder." It asks before deleting, sending, buying or writing files, and "rm, sudo, disk tools and shutdown are always blocked." A "Watch for me" feature stares at a window or web page until something happens, and "Coding agents like Codex are watched through their own log, with no screen at all." A Meeting Copilot overlay keeps transcripts on the Mac and never saves audio.

He answers the thread like someone who has shipped before. On local models: "Smaller models, like Haiku, I have found to already struggle during multi-step tasks," followed by a custom server option for DeepSeek, Ollama and LM Studio anyway. A reader ported it to Windows during the thread; it is now linked in the README. Credits go to Plip for some of the logic and to Jakub Antalik's thinking-orbs for the animation.

Why we picked it

Because we are building a desk assistant on Claude right now, and a 9th grader shipped one for the whole Mac, with the right defaults and a sandbox.

And because of the learning loop. Most agents want more of your tokens. This one saves the steps and tries never to ask the model the same thing twice. That is the quietest kind of rate limiting, and the only kind we endorse.

sources: Hacker News · source code · docs · demo · commit history

open the article →

Franz Enzenhofer's agent could refactor a monorepo but could not tell him which "Allow" to click.

score AI inside: no

  • bigarrow is a macOS command line tool that draws a big arrow and a sign over whatever is on your screen.
  • Install is two lines, brew install franzenzenhofer/tap/bigarrow and bigarrow install-skill.
  • Franz Enzenhofer is an SEO consultant in Vienna, and the project went from first release to version 0.4.5 in two days, October 8 and 9, with a round of fixes after the Hacker News feedback.
  • He credits Peter Steinberger's Peekaboo and Nameplate for the overlay recipe and the skill packaging.
  • The agent did the work and now needs you to click Allow, and you cannot find Allow.

open the article →

What it is

bigarrow is a macOS command line tool that draws a big arrow and a sign over whatever is on your screen. Clicks go through it, your keyboard focus stays where it was, and the arrow deletes itself when its time is up. It ships with a skill for Claude Code and Codex, so the agent knows when to point and how. The README's pitch: "Your AI agent can refactor a monorepo, write a migration and explain monads." Then: "bigarrow gives it a finger."

The use case is every dialog an agent cannot click for you. In the Hacker News thread, Franz Enzenhofer explained: "Claude refuses to do certain actions (enter passwords, change security settings, create new accounts on external services)." So instead of a paragraph in the terminal, the agent runs one command and a sign appears next to the actual button: "Franz, click Allow." Values to paste become a copy button on the sign, and --say reads it aloud for anyone not watching.

Install is two lines, brew install franzenzenhofer/tap/bigarrow and bigarrow install-skill. Arrows come in bend, straight, zigzag and spiral. The README's demo scenes include "Delete node_modules? Yes. Obviously." and a Friday deploy where "the agent votes Cancel." The FAQ answers "Is it AI?" with "No. It is the least intelligent part of your AI stack, and proud of it."

How it was built

Franz Enzenhofer is an SEO consultant in Vienna, and the project went from first release to version 0.4.5 in two days, October 8 and 9, with a round of fixes after the Hacker News feedback. The commit history lists Claude alongside him on commit after commit, and the repo carries a CLAUDE.md that reads like a shop manual: Swift 6 strict concurrency, 200 lines per file, SwiftLint in strict mode before any commit, no mocks. The agent is also told where it may not draw. On Franz's work Mac it only builds, runs unit tests and renders arrows offscreen. Screen tests run on CI and on a second Mac over SSH.

The look took arguing. "i did spare quite a lot with it of how i want the arrows to look like," he told one commenter. "the first versions looked way to professional and lost that 'yes, that, exactly that, i mean it' arrow feeling." The README counts 104 automated tests and 18 behaviour checks, and reports the skill costs the agent "182 tokens, most of the time." The hardest bug, per the FAQ, was keeping the arrow on top without taking focus: "On top and focused are two different things on macOS."

He credits Peter Steinberger's Peekaboo and Nameplate for the overlay recipe and the skill packaging.

Why we picked it

Because the gap it fills is exactly where our brand lives. The agent did the work and now needs you to click Allow, and you cannot find Allow. A tool whose whole job is to never click, type or capture, and to be proud of how little it knows, is a good neighbour for a shirt about permissions.

And because of the detail in the project guide: the agent that drew every arrow in the README is forbidden from drawing a single one on its maker's own machine.

sources: Hacker News · source code · docs · demo · commit history

open the article →

Anup Ray bought a Nokia 110 4G to cut his screen time.

score RAM: 48MB

  • A Nokia 110 4G with an AI agent in it.
  • Anup Ray first tried the obvious thing.
  • The loop itself is plain: "It sends your message to DeepSeek chat API and use tool calls accordingly."
  • Status, in his words: "It's a working prototype loaded into RAM from a computer."
  • No install path, no room for a normal agent, every feature tested one at a time on the handset.

open the article →

What it is

A Nokia 110 4G with an AI agent in it. You open what used to be the Calculator, type on the number pad, and the agent answers over the phone's SIM data. Ask it to and it will check the battery percentage, turn the torch on or off, start a phone call or set an alarm. The demo video is 48MB of phone doing things 48MB of phone was not supposed to do.

The origin story is the best part. In his Show HN: "I was using this mobile to reduce my screentime. Recently got the idea to put ai agent in it."

How it was built

Anup Ray first tried the obvious thing. "Started with installing android, but failed as it has only 48MB RAM." So he reverse-engineered the firmware instead, "for few days." It is written in C, he told the thread, unlike the Java of earlier Nokias, so custom changes are possible. They just get erased on restart.

His README lays it out in five phases. Phase 1 was a workaround: an AI search option bolted onto Opera Mini. Phase 2 was getting his own code running at all. With no supported way to install anything, he loads custom code into RAM through the Calculator's entry point. The first proof was a tiny C app that printed "PDP" on keypresses. Phase 3 was a real answer over SIM data, which meant tracing firmware callbacks until a response reached the app with the USB cable unplugged. Phase 4 made the chat usable, swapping the calculator layout for a text input; a colour change kept rebooting the phone until he found he was writing the wrong style field. Phase 5 turned chat into actions. "As the phone memory is only 48MB, normal agent code wont fit," so he wrote a minimal custom agent, shrank the builds, compressed the payloads and reused buffers.

The loop itself is plain: "It sends your message to DeepSeek chat API and use tool calls accordingly." And the coding help: "I used OpenAI Codex to help build it and tested the app on the handset."

Status, in his words: "It's a working prototype loaded into RAM from a computer." Once loaded it runs unplugged on mobile data. Power it off and it needs loading again. "Permanent installation is still unresolved." The source is not published yet, because "sensitive data is scattered across the project."

Why we picked it

Because he bought the one phone that cannot run an app and then taught it to run an agent. The screen-time plan did not survive contact with the idea, and we respect that.

And because the constraint is the whole story. No install path, no room for a normal agent, every feature tested one at a time on the handset. As one commenter put it, finding a Nokia 110 is the hard part.

sources: Hacker News · source code · demo

open the article →

Einar Andersson gave Opus 5.5 one prompt and a zip of the original SimCity 2000.

score subagents: 27

  • City 2026 is SimCity 2000 in a browser tab, drawn low-poly.
  • Einar Andersson is the technical co-founder of SeenThis, in Sweden, and the person behind dator.dev, a page of side projects like a browser Quake III map editor.
  • The whole story is in his Hacker News post.
  • After that, "it has basically just been polishing."
  • It is that, asked to be accurate, it reached for an emulator and a 10,000-day replay instead of a shrug.

open the article →

What it is

City 2026 is SimCity 2000 in a browser tab, drawn low-poly. The page says "Loading…" and then "Click to start," and that is the whole front door. Behind it is, in the maker's words, "a clone of simcity 2000 I (using Opus 5.5) built over a few days." A "full replica of game logic," not a tribute. The original rules, zoning, budgets and disasters, running in WebGL.

He added a few things the original did not have: smog, depth of field and environmental sounds, "to make it a bit more immersive without killing the original nostalgic vibe." If it runs slow, depth of field is on by default and Options > Display turns it off. That was his one reply in the thread.

How it was built

Einar Andersson is the technical co-founder of SeenThis, in Sweden, and the person behind dator.dev, a page of side projects like a browser Quake III map editor. His GitHub bio: "thank you, computer."

The whole story is in his Hacker News post. It started with a single prompt: "create a typescript webgl low-poly version (full replica of game logic) of simcity 2000," with a zip of the original game attached for inspection. Opus "decompiled it using Ghidra and started translating the old native PowerPC code function-by-function." Then: "After about 2 hours, using 27 subagents, i had a running version in my browser."

The verification step is our favorite part. He followed up with "make sure all aspects of the game is implemented correctly and accurately," and the agent chose its own method: it picked up the Unicorn PowerPC emulator, wrote "a harness with faked Mac system calls in Python," and ran the original and the port side by side for 10,000 game days until they produced "an identical game state."

After that, "it has basically just been polishing." The 3D and vector assets are generated with code by Opus, the bitmaps by GPT Image 2. The music and sounds are not AI at all, "because all AI audio models i've tried sound terrible." They are CC-licensed, with music by Kevin MacLeod.

Why we picked it

Because the surprising part is not that an agent wrote a city builder. It is that, asked to be accurate, it reached for an emulator and a 10,000-day replay instead of a shrug. That is a test plan most of us would not have written for ourselves.

And because of the reason he gave: "I created this because I wanted it and to see how capable Opus 5.5 was." Thank you, computer.

sources: Hacker News · demo · dator.dev

open the article →

Jason Tragakis pulled 526 painters and 82,139 works out of Wikipedia and built a museum you walk with WASD, from Li Cheng to Banksy. The 3D gallery started, by his own account, as a single prompt.

score paintings hung: 82,139

  • A Walkable History of Art is a museum in a browser tab.
  • Jason Tragakis, of justdataplease.com, put it on Hacker News today.
  • The AI part comes from the thread.
  • Also, someone complained they could not jump onto the benches.
  • He is Greek, and says Greek history might be the next building.

open the article →

What it is

A Walkable History of Art is a museum in a browser tab. The tagline on the door: "Every artist, every work, under one roof." You scroll a timeline across more than a thousand years of painting, click a period, pick a painter, and walk into their gallery with WASD. Hold W to sprint. On a phone you drag to look and tap the floor to walk.

The collection comes from Wikipedia, Wikidata and Wikimedia Commons: 526 artists and 82,139 artworks across 35 periods and traditions, Chinese, Japanese, Indian and Persian as well as European. The rooms are arranged by date, each period gets its own room styling and music, and clicking a painting shows its details, dimensions and Wikipedia source. The Hacker News title sums up the range: "From Li Cheng to Banksy."

How it was built

Jason Tragakis, of justdataplease.com, put it on Hacker News today. The repo is MIT and the development guide is specific about the engineering: Next.js 16 with every one of the 526 galleries prerendered, React Three Fiber galleries that only render when something changes so "standing still costs no frames," and a fixed pool of 16 spotlights handed to whichever works are in view "so walking never recompiles a shader."

The AI part comes from the thread. A commenter said they had built almost the same 3D gallery with Claude last year, down to the lights, the benches and the walls. Jason's reply: "it started as single prompt so the model did its own work," and then, "after that i worked a lot on this." The ingest pipeline, the repair scripts and the performance probe are the "worked a lot" part.

Why we picked it

Because it is the best kind of project: a data person pointed a model at a museum and then did the real work of making 82,139 Wikipedia images walk at a decent frame rate. The footer gives all credit to Wikipedia and asks you to donate, the right instinct for a thing built on free knowledge.

Also, someone complained they could not jump onto the benches. Jason's answer: press P and enter the password Vandalism0. "This is supposed to be for educational purposes," he wrote, "but since you want a bit of fun."

He is Greek, and says Greek history might be the next building. We will take a ticket.

sources: Hacker News · source code · docs · demo · justdataplease.com

open the article →

October 6, 2026 · built by Péter Szentkirályi

Darkplug: an iPhone and a $20 smart plug become a darkroom f-stop timer

Péter Szentkirályi wanted a proper f-stop timer for darkroom printing and didn't want to spend a couple hundred dollars on one.

score smart plug: $20

  • Darkplug is, in its maker's words, "a darkroom timer for iPhone that controls your enlarger."
  • It does regular timing and f-stop timing, with a four second count-in option.
  • The post's title puts the switch at $20.
  • The hard part was never the code.
  • The first question in the thread was about network jitter between the tap and the plug switching.

open the article →

What it is

Darkplug is, in its maker's words, "a darkroom timer for iPhone that controls your enlarger." You plug the enlarger into a Shelly smart plug, put the plug and the phone on the same network, type the plug's local IP address into the app, and the phone switches the lamp on and off.

That is the whole trick. The app does the timing, which phones are already good at, and the plug does the one thing a phone cannot: switch a mains lamp. The interface is red on black so it is safe under a safelight.

It does regular timing and f-stop timing, with a four second count-in option. The paid Darkplug+ tier (a one-time purchase, no subscription) adds test strips, a dodge and burn planner with a grid, split-grade printing with two timing channels, and a paper development timer. The free app is on the App Store for iPhone and iPad.

How it was built

Péter is a film photographer who does some darkroom printing. On Hacker News he says he "always wanted one of the proper, advanced f-stop timers, but never really wanted to spend a couple hundred dollars on one." Then: "Eventually I figured my phone was already perfectly capable of doing the timing, it just needed a way to switch the enlarger."

The post's title puts the switch at $20. The build method gets one line, and it is our favorite line of the week: "Fair warning: the app was blindly vibe coded using Codex and Xcode."

Blindly, and then tested the honest way. He has been using it "for a few months in my improvised bathroom darkroom."

Why we picked it

Because this is the vibe coding story we like best: not a developer shipping another dev tool, but a photographer with a specific, slightly expensive problem who could suddenly build the fix himself. The hard part was never the code. It was knowing what a printer needs from a timer, and a person who prints knows that.

It also ends in a physical thing. Most agent projects live in a browser tab. This one lives in a bathroom with the lights off, switching a lamp while someone counts in f-stops.

The first question in the thread was about network jitter between the tap and the plug switching. Péter's answer, in effect, is a few months of prints. It works on his prompt.

sources: Hacker News · docs · demo

open the article →

Alice built an oil-paint simulator in Rust and handed the brushes to Claude, GPT, Gemini and friends. No image model anywhere.

  • stillwet is a gallery of oil paintings made by AI models, and none of them came out of an image generator.
  • Underneath is claude-paint, which the README describes as "a physical oil-paint simulator in Rust, and an easel to paint with it."
  • Alice runs it in rounds, each with a different setup, and writes up every one.
  • Asked to plan a painting with no studio context, the site reports, "Claude Opus chose a jug with lemons six times out of six."
  • The code is MIT, and the paintings, logs and texts are CC BY 4.0.

open the article →

What it is

stillwet is a gallery of oil paintings made by AI models, and none of them came out of an image generator. In the site's words: "Each one paints by writing a program against a simulation of oil paint on linen: bristle brushes, wet paint, drying, layered glazes. No image model is involved."

Most of the painters work after Caspar David Friedrich, the German Romantic, "from written research alone; they never see a picture of his work." Others get a free subject. There are 75 paintings up so far, from Claude Opus 5.5 (the most prolific), several GPT-6 models, Gemini, Kimi, DeepSeek and more, plus a live studio where you can watch painters at work.

How it was built

Underneath is claude-paint, which the README describes as "a physical oil-paint simulator in Rust, and an easel to paint with it." Simulated bristles carry wet paint over a primed linen canvas, the paint levels and dries on a clock, and layers combine by Kubelka–Munk optics, the same math used to describe real pigment.

The easel is a live Lua session. Some painters write the whole picture as one program, but 46 of the 75 were painted the slow way: a passage at a time, stepping back to look. Since round 19 the painters get only the easel's own tools: paint, look, keep a journal and read the studio's notes.

Alice runs it in rounds, each with a different setup, and writes up every one. The repo's tagline is "Teaching claudes to paint", and she says on Hacker News that she let Claude Opus 5.5 pick the name, "since they had such a big part in it." The proper write-up is still to come. Until then, her advice is to "point your favorite agent at the git history."

Why we picked it

Because the findings page is as good as the paintings. Asked to plan a painting with no studio context, the site reports, "Claude Opus chose a jug with lemons six times out of six." Two winter painters, who never saw each other's work, gave their pictures the identical title: "Hünengrab im Schnee am Abend (Dolmen in Snow at Evening)."

It is also a different shape of agent project. Nobody asked for an app. Someone built a careful physical simulation, gave the models brushes instead of a pixel grid, and then paid attention to what they did with them. The thread reached the Hacker News front page, and Alice, by her own account, didn't notice until it was already there.

The code is MIT, and the paintings, logs and texts are CC BY 4.0. Start with the gallery, then read the rounds.

sources: stillwet.art · source code · docs · demo · the rounds

open the article →

anteloc has spent since last December teaching ChatGPT and Claude to write LDraw, the "assembly language" of LEGO CAD. The result is an open-source tool where an agent builds the model brick by brick.

  • ldraw-nova is an open-source toolset that lets an AI agent design LEGO models.
  • It ships as a Docker web app that works with OpenAI, Claude or OpenRouter.
  • a 3D viewer, a 3D player, and a VR view for Meta Quest 3
  • That last one is the fun part.
  • It was the maker's first post on Hacker News, and it picked up more than 130 points.

open the article →

What it is

ldraw-nova is an open-source toolset that lets an AI agent design LEGO models. You describe what you want, guide it if you feel like it, and the agent writes the model in LDraw, a plain-text format where every line places one part at one position and rotation. Open the file in LDView, LeoCAD or Studio and it is a real LEGO CAD model you can spin around and edit.

The maker calls LDraw an assembly language, pun intended, and the pitch follows from that: if agents can write code, and a LEGO model is code, then agents can build LEGO.

How it was built

By the maker's own account this started last December as an experiment: could ChatGPT and Claude write LDraw the way they write Python? LDraw is almost all positions and rotations, the kind of math models are usually bad at, so the surprise was that early tests went well. Months of iteration later, using GPT-6 Astra and Opus 5.5, it has turned into a Python toolset plus instructions and docs written for agents: part search, example lookup, and the rules of the format.

It ships as a Docker web app that works with OpenAI, Claude or OpenRouter. When a build finishes you get:

  • the LDraw source for the model
  • a 3D viewer, a 3D player, and a VR view for Meta Quest 3
  • a Blender-editable .glb file
  • the full chat history and the agent's thinking

That last one is the fun part. Each generated file records what the agent was picturing before it placed a single brick.

Why we picked it

Most agent demos end in a web app. This one ends in something you could put on a shelf. It is also a good example of what the work actually looks like: the model didn't change, the tooling around it did. Give an agent a searchable parts catalog and a format it can check, and "bad at math" stops being the blocker.

It was the maker's first post on Hacker News, and it picked up more than 130 points. The project is AGPL-3.0 and there is a gallery of generated models and a demo video if you want to see results before you install anything.

sources: Hacker News · source code · docs · demo · demo video

open the article →

Built something? Tell us at [email protected] or tag @AcceptAllHQ. Weekly news lives in the Vibe Log.