One Android agent picked up 5,300 GitHub stars this week. Somebody spent three months building an MMO and called it quits. Somebody else put all the world's law into Git. And a guy turned his AI assistant into a Tamagotchi, which 1,376 people upvoted.
The lead story is games, and it's an honest one: the AI gets you a working game fast, and finishing it is still the hard part. Elsewhere, a cheap model does the typing while an expensive one does the thinking, two people built nearly the same Hacker News book index without knowing it, and someone ran Claude on a terminal from the 1980s.
New this issue: Back on the radar, where we check in on builders from last week who shipped again. The meteor camera did not slow down. Neither did the guy tracking whether Opus is getting dumber. (It isn't. Day 34.)
01
Lead story
AI will build your game. Finishing it is still on you.
The most-discussed builds this week were games, and the comment sections read like a support group. A solo developer in Korea spent five months building a WW2 tank strategy game with Claude Code; it works, and the post is about what comes after it works. Another builder spent three months vibe-coding an MMO and announced they were throwing in the towel. Over 360 people weighed in on each.
The biggest argument of the week came from a tabletop game master who built a full game system for Foundry VTT with Claude Code and was told that counts as stealing. That thread drew 438 comments.
On the brighter side: a fishing and exploration game reached a free demo after three months of AI-only coding, browser games that turn several phones pushed together into one big board picked up 2,295 upvotes, and someone finally shipped the Earth-scale explorable planet they shelved in 2012. The pattern is clear. AI gets you a playable game in weeks. Making it fun, finished and fair is still on you.
Games & hobbiesClaude Code
Five months, one dev, a WW2 tank war
A solo developer in Korea shipped a working WW2 tank strategy game with Claude Code. The post is about the part nobody warns you about: what comes after it works.
A game master built a full tabletop system for Foundry VTT with Claude Code and was told that counts as stealing. 438 comments later, nobody agreed on anything.
Each of these is good on its own. Bolt two together and you get something nobody has shipped yet. The builders have never met. Consider this the introduction.
bot_boardOn its own: lets a crew of agents coordinate on a shared board, but someone still has to watch it.
+
agentringOn its own: calls your phone when one Claude Code session needs you.
→
TogetherA team of coding agents that coordinate on their own and phone you only when they're stuck
The team room and the phone call
One builder replaced their agent orchestrator with a plain message board where coding agents post, read and coordinate. Another made Claude Code call their phone when it needs input, with no Twilio and no phone number.
How it fits: The message board as the agents' shared room, with agentring wired in so a stuck agent rings the human instead of waiting in silence.
floorplanOn its own: draws an accurate map of your codebase, but it sits still.
+
ProschiOn its own: animates requests through a system diagram, but you draw the diagram by hand.
→
TogetherA live architecture diagram of your real codebase, with requests moving through it
Map the code, then watch it run
Floorplan has Claude or Codex draw a plain-English, interactive map of how your app fits together. Proschi is a text language for system diagrams that plays simulated request flows right in the terminal.
How it fits: Floorplan's map exported as a Proschi diagram, so the drawing comes from the real code and the request flows show how it actually behaves.
cheapAgOn its own: cuts the expensive model's share of the work, without showing where the money went.
+
optimAIzrOn its own: shows where your usage goes across tools, without changing who does the work.
→
TogetherSee exactly where your AI coding budget goes, then send the cheap work to a cheap model
Track the spend, then cut it
cheapAg keeps Opus as the planner and hands boilerplate to a model that costs a fraction of a cent, because its builder kept hitting Opus limits. optimAIzr tracks AI coding usage across Claude Code, Codex and opencode, now with a VS Code extension.
How it fits: optimAIzr spotting which tasks burn the most tokens, and cheapAg routing exactly those to the cheaper model.
CrowdGPTOn its own: trains a model with volunteers' hardware, but still needs somewhere to run it.
+
ClusterChatOn its own: serves a big model on volunteers' GPUs, but it's someone else's model.
→
TogetherAn AI model trained by volunteers and served by volunteers, with no data center at all
A people-powered model, end to end
CrowdGPT, now on version 2.6, trains a language model on volunteers' computers. ClusterChat runs a 27B model in the browser by splitting it across volunteers' consumer GPUs.
How it fits: CrowdGPT's community-trained model served through ClusterChat's volunteer network: a model the public both builds and runs.
03
Parallel builders
Same idea, different builders
Two iPhone agents, two Hacker News indexes, two ways to squeeze a 125B model onto cheap hardware. Same problem, different builders, nobody comparing notes. Somebody start a group chat.
Thumbwise is a Mac app plus MCP server that lets Claude rearrange your iPhone Home Screen over USB. OpenLolo is an open-source hardware box that lets AI assistants see and operate an iPhone through MCP.
If they talked: One is software, one is hardware, and they're solving the same problem. Sharing an MCP interface would let an assistant work with either.
HackerBooks indexes about 16,500 books mentioned in Hacker News comments, each linked back to the discussion. Days later, HackerHack launched as the same idea for posts shared on HN.
If they talked: These could be one site with two tabs. Right now they're two sites with one audience.
Betelgeuse runs Qwen 3.8 Flash Next at 22 tokens a second on a 12GB RTX 3060 by predicting which experts it needs. MoEShift runs the same 125B model on a 24GB Mac and nearly quadrupled its speed.
If they talked: Same model, same trick of being clever about which parts to load, one on Nvidia and one on Apple. They'd learn a lot from each other's code.
One builder used Claude to put all of the world's law into Git, where every change is tracked. Another built an interactive explorer of 344,799 cross-references between Bible verses.
If they talked: Both turn an enormous body of text into something you can navigate and follow the connections in. The tooling for one would serve the other.
04
Five that got people talking
The ones everyone argued about
1
Developer tools & agentsLLM
Mika
An agent that lives on your Android phone and taps through apps and websites for you. 5,300 stars and a thousand comments in seven days.
Builders from issue two who came back with news. The meteor camera nearly tripled its upvotes, the volunteer-trained model got a version bump, and the Opus-nerf tracker is still finding nothing, which is the point.
Hardware & homeClaude Code
openskycam
The meteor camera from issue 2 posted again and went from 602 to 1,684 upvotes. Claude Code still sorts the night's catch before the builder wakes up.
Last issue: Open-source 4K sky camera that Claude Code runs to detect and sort meteors.
793 builds, 374 audiences. Developers building for developers, sure. The good stuff is down the long tail: birders, DJs, long-distance couples, two kids sharing one iPad. Tap one and the Explorer serves up everything made for them.
07
Under the radar
Big in a small room
Appetizers before the main course. Each of these was the biggest thing in its own corner of the internet and never came close to the front page. Hardly an upvote between them, but the comment threads ran long. A stranger who stops to argue is worth ten who scroll past.
Business & work
Knockbox
Script widget that measures demand for a feature before you build it.
Still half the radar. This week: an agent harness that actually gets its pull requests merged, memory that fits in a gigabyte, a diff that reads PDFs, and a local model with its own phone number.
Developer tools & agentsLLM
DevBox
Hands Linear tickets to coding agents and ships the result. The builder's whole team adopted it, and about 80% of its pull requests get merged.
