This is the first of a weekly list I plan to keep: a handful of open-source repositories that caught my eye, described only from what their READMEs say, with the numbers I could check myself.
How I picked them: on 2026-11-06 (UTC) I queried the GitHub search API for repositories created in the last 30 days with more than 300 stars, sorted by stars. I also tried GitHub's weekly trending page and Show HN on Hacker News. Neither worked from the sandbox that runs this job (the trending page returned 403 and the Hacker News API connection was refused), so there are no "stars this week" figures here. Every number below is cumulative and was read on 2026-11-06.
From the top 70 of 753 matches I skipped game recompilations and emulators of commercial titles, Android TV debloaters, prompt collections and anything without a license. What was left and useful to people who build web apps, desktop apps or games came down to six.
pixelfx: 244 pixel art effects as code
pixelfx is a library of pixel art effects (fire, explosions, weather, UI pickups, screen transitions, critters) where every effect is a small JavaScript file under src/fx/<category>/. The README shows two ways to use it: import the code directly into a canvas, Phaser, PixiJS, Kaplay or three.js game with no dependencies, or run node tools/render.mjs <effect> --sheet to export a sprite sheet plus an Aseprite-style JSON for Godot, Unity or other engines. Effects accept a palette, so a fire can be recolored without editing its code. It also ships a Claude Code skill for making new effects.
Signals: 5,585 stars, 734 forks, MIT, created 2026-10-30, last push 2026-10-31. No tags yet. Install is npm i github:patch-gg/pixelfx, not a registry package.
Try it if you are prototyping a browser game and need placeholder or even final VFX. Skip it if your game isn't pixel art. Caveat: the README says some files adapt code from Hugo Duprez's effects at Sprite Fusion. The repo license is MIT, but check the header of any effect you ship commercially.
OpenIDEA: an IntelliJ-style IDE on a Rust core
OpenIDEA brings the IntelliJ IDEA keymap, tool windows and search-everywhere habit to a native app. The README lists an editor, code insight, debugger, Git, terminal, database and HTTP clients for 60+ languages, on macOS, Windows and Linux. Plugins run as sandboxed WebAssembly components, and the README calls the plugin API a preview.
The README includes a benchmark opening freqtrade (about 770 Python files): 1.1 s to ready versus 13.2 s for IntelliJ IDEA 2026.2.3, and 169 MB versus 2011 MB of memory. Those are the project's own measurements, and the README itself notes that IDEA analyses Python more deeply, so "ready" is not the same amount of work. The perf/compare folder lets you rerun it.
Signals: 7,425 stars, 620 forks, Apache-2.0 (fine for commercial use), created 2026-10-21, last push 2026-11-06. Latest tag v0.1.38. GitHub's language detector reports TypeScript, which fits a Tauri shell around the Rust core (there is a src-tauri folder).
Try it if JetBrains muscle memory is the main thing keeping you on a heavy IDE. Skip it if you rely on deep framework-specific inspections today. Caveat: it is three weeks old and not affiliated with JetBrains.
Floorplan: a map of an app you built with an agent
Floorplan is a Claude Code and Codex plugin that reads your codebase and draws an interactive, zoomable map: parts, files, endpoints, screens, and which frontend request hits which server endpoint, even across repos. It labels things in plain English, and has a "Recent work" view that highlights what the agent just changed. The part I find most practical is the health check, which looks for keys in code, Supabase tables anyone can read, open Firebase rules, Stripe webhooks that could be faked and endpoints with no sign-in. The README calls it "a check, not an audit", and I agree with that framing.
Signals: 3,076 stars, 234 forks, MIT, created 2026-10-27, last push 2026-11-02. Latest tag v0.6.3. Node 18+, zero dependencies, runs on your existing agent plan with no API key, and nothing is committed to git.
Try it if you inherited or vibe-coded a Next.js plus Supabase app and want a quick picture and a first security pass. Skip it if you already maintain architecture docs. Caveat: the map is drawn by a model reading your code, so verify anything you act on.
unslop: a design pass for AI-looking UI
unslop is a skill for Claude Code and Codex. It audits a screen for what its README calls slop (purple gradients, glows, emoji labels, shouty copy, uneven spacing), proposes a direction based on your product, then fixes it in passes: spacing, color, icons, copy, type, layout, details, visuals, motion. It proposes before editing, and asks before any paid image generation.
Signals: 10,701 stars, 584 forks, created 2026-10-29, last push 2026-11-01. The license is CC BY 4.0, not a software license. The README says commercial use is allowed if you credit the author and link the repo. That is easy for a team tool, but read it before you vendor the skill into a product.
Try it if your dashboard looks like every other generated dashboard. Skip it if you have a design system and a designer. Caveat: the README also advertises a paid service by the same author.
Rune: a GPU-rendered UI framework in Rust
Rune is a native UI framework with its own renderer, text, layout and spring animations, targeting macOS, Windows, Linux and the web. Every app can run headless, so a coding agent can render it to PNG, dump the element tree and drive it with scripted input. That is a sensible design for agent-built UI.
Signals: 2,458 stars, 58 forks, MIT, created 2026-10-22, last push 2026-11-06. No tags. The README says it is pre-release, asks you to pin a git revision of dev, and states that the project does not accept pull requests (issues only).
Try it if you are building a Rust desktop tool and want to test UI without a display. Skip it for anything you need to ship this quarter.
TokenEconomizer: cheaper coding-agent sessions
TokenEconomizer is a research plugin from Microsoft Research, Microsoft Security and UC Berkeley for Claude Code, Codex, GitHub Copilot CLI and Pi, with experimental support for others. It learns from past sessions in a repo: it remembers facts like "use python -m pytest here", stops loading tools that go unused, and redirects a turn once when a judge model thinks it is stuck. The paper reports 18.4–30.9% lower mean cost on SWE-Bench tasks across five agent and model combinations. That is their measurement, not mine.
Signals: 4,893 stars, 299 forks, Apache-2.0, created 2026-10-13, last push 2026-11-06. Latest tag v0.1.0. Data stays in ~/.tokeneconomizer.
Try it if you run the same agent on the same repo every day. Skip it for one-off repos, since it needs sessions to learn from. Caveat: tool pruning edits your agent's settings. The README says one command restores a tool.
Side by side
| Project | Category | Stars (2026-11-06) | License | Latest tag | Maturity |
|---|---|---|---|---|---|
| pixelfx | Game VFX | 5,585 | MIT | none | days old |
| OpenIDEA | Desktop IDE | 7,425 | Apache-2.0 | v0.1.38 | early releases |
| Floorplan | Codebase map | 3,076 | MIT | v0.6.3 | early releases |
| unslop | UI design skill | 10,701 | CC BY 4.0 | none | days old |
| Rune | Rust UI framework | 2,458 | MIT | none | pre-release |
| TokenEconomizer | Agent cost control | 4,893 | Apache-2.0 | v0.1.0 | research code |
Takeaway
Five of the six involve coding agents somehow, as plugins, skills or tools designed for agents to drive, and that matches what the top of the last 30 days of GitHub search looked like. If I had to try one this week it would be Floorplan's health check, because it costs nothing and looks for mistakes that are expensive later. All six are less than six weeks old, so star counts here measure attention, not reliability.
Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.
Top comments (0)