Focus: The GitHub repos gaining the most stars this week — momentum and breakouts across open source.
Method: Ranked by stars gained over the past week, scraped from GitHub's weekly trending pages.
Overview
One repo ran away with the week: morluto/rea's +214,837 stars is roughly eight times the number-two entry, and reverse-engineering tooling broadly had a strong showing alongside it, with NSA's Ghidra and the penetration-testing agent PentAGI both posting real weekly gains despite being established, years-old projects. The rest of the board is thick with the Claude Code and Codex skills ecosystem — packages from Matt Pocock, Addy Osmani, Vercel Labs and others, mostly distributed through the same npx skills add install flow. Video tooling for agents showed up three times (hyperframes, OpenMontage, Remotion), and two repeats from past editions of this list — hyperframes and free-claude-code — held almost exactly their prior pace rather than fading. Not everything here is AI: a self-hosted gym tracker and a lyrics-animation music player both out-gained more than half the AI entries on the board.
REA hooks an AI agent up to decompilers, disassemblers and app inspectors so it can explain what a piece of software actually does, down to the x86 instructions, without ever touching the source. The README's pitch is blunt: see a feature in someone else's app, point an agent at it with REA, and get a working explanation of how it was built. A quarter-million stars in a week is an outlier even for this list, and the case studies the README leads with — recovering game logic from old DOS binaries, reconstructing an Electron clipboard bridge — read like the social proof that tipped it over.
AnyPS5 relinks PS5 executables into native Windows and Linux binaries — no emulator, no separate runtime process, just a relinker that rewrites the binary to call host system libraries directly. Porting at the executable level instead of emulating the whole console is a much harder trick than it sounds, and that distinction is probably what's pulling in people who'd otherwise scroll past this as just another emulator.
ArtCraft is a visual IDE for blocking out a scene in 2D or 3D, posing characters, and then handing the final render to one of 62 supported AI models from FLUX, Seedream, Kling and others. The idea is to keep a human directing composition while generative models do the rendering, rather than typing a prompt and hoping. A two-minute demo video linked straight from the README is doing a lot of the work here — visual tools tend to spread through a screenshot before anyone reads a changelog.
Matt Pocock, who runs the Total TypeScript and AI Hero newsletters for roughly 60,000 subscribers, packaged the agent skills he uses on his own engineering work into a one-command install for Claude Code, Codex and GitHub Copilot. The pitch pushes back on heavier spec-driven frameworks like GSD, BMAD and Spec-Kit, which he argues take control away from the engineer the moment something goes wrong. With an audience already primed to try whatever he ships, 292,000 total stars looks like a newsletter list cashing in an install link more than organic discovery.
This is an agent skill rather than an app: point Claude or another coding agent at it and it draws diagrams as a single self-contained HTML and SVG file, across 45 layouts, matched to your brand colors. The README's own jab — "no shadows, no Mermaid slop" — is aimed at the washed-out auto-generated diagrams coding agents tend to produce by default. It installs with one npx skills add command, the same lightweight distribution model Vercel Labs' skills CLI further down this board is built around.
Moli is a headless browser written in Rust, built specifically for AI agents to drive, under the tagline "structure first, pixels on demand" — an agent gets the page's DOM and structure by default and only pays for rendered pixels when something actually needs a screenshot. That's a different design target than Playwright or Puppeteer, which were built for humans watching a browser render and only later picked up agent wrappers.
iFixAi audits what an AI agent actually did against what it was supposed to do, catching the mistakes and blind spots an agent's own summary tends to paper over. It landed as the top Product Hunt post the day it launched and was Trendshift's #1 Python repository of the week — about as clean a traceable cause as a star spike gets on this list, a coordinated launch rather than a slow organic climb.
Ghidra is the NSA's own reverse-engineering suite — disassembler, decompiler, scripting, the lot — open source since 2019, so a project this established doesn't usually show up on a weekly-gain chart. Its appearance here likely rides the same wave as REA and PentAGI elsewhere on this board: AI agents are suddenly good at driving tools like this one, and people tracing the newer agent-wrapped options back to the heavyweight original would explain an old project posting a fresh week like this.
openGym is a self-hosted workout and body-weight tracker: plan a week of training, log sets through a guided workout screen, sync across your own devices behind passkey login, no account on anyone else's server. It's one docker compose up to run, and it ships a live in-browser demo so you can try the real app before installing anything. The appeal is squarely about owning your data rather than anything AI-related — a useful reminder that this list has no topic filter.
Hyperframes takes HTML as input and renders it out as a video file, aimed at agents that can write markup easily but have no native way to produce something a video player opens. Built by HeyGen, the AI video company, it's a repeat from this list's September 15 edition, where it posted +5,052 stars — this week's +5,123 means the climb hasn't slowed in a month.
Addy Osmani, a longtime Google Chrome engineer known for writing about web performance, packaged the skills and slash commands he uses to take a project from spec through build, test, QA and ship with a coding agent. Nine commands map to that pipeline, each pulling in whichever skill the phase needs. An existing following plus a Trendshift badge on day one is a familiar pattern by now: a well-known engineer ships a skills package and the audience converts fast.
cc-switch is a desktop app for juggling the dozen-plus coding-agent CLIs now in circulation — Claude Code, Codex, Gemini CLI, OpenCode and others — letting you switch API providers and manage MCP servers, skills and prompts from one place instead of hand-editing config files for each tool. At 144,619 total stars it's already one of the larger projects on this week's board, and the swelling number of competing agent CLIs it manages is probably why a switcher like this keeps finding new users.
PentAGI runs autonomous agents through penetration-testing workflows — reconnaissance, exploitation, reporting — under what it calls agent supervision, with a Discord and Telegram community built up around it. It's in good company this week: between PentAGI, Ghidra and REA, offensive and reverse-engineering tooling is having a stronger week than usual on this board.
free-claude-code is a pass-through that routes Claude Code, Codex, Pi, OpenCode and several other coding-agent harnesses against a shared pool the author says totals more than 1.3 billion free tokens. It's a repeat from this list's August 25 edition, where it gained +3,877 stars; this week's +3,764 is close enough that the climb looks flat rather than fading. The pitch is entirely about the cost of running these agents rather than their capability, so it keeps spreading fastest among people who already use the tools it fronts for.
FreeLLMAPI aggregates the free tiers of 34 LLM providers — 635 model endpoints in total — behind one OpenAI-compatible /v1 API, with a router that picks a live model per request and fails over the moment a provider hits its rate limit. It ships as a self-hosted Docker image alongside native macOS, Windows, iOS and Android apps, a wider release surface than most entries on this board manage.
Also Climbing
tester-army/e2e — +3,505 this week — an open source end-to-end testing framework for web and mobile that hit GitHub's trending-repo-of-the-day badge, built by TesterArmy.
calesthio/OpenMontage — +3,387 this week — bills itself as the first open-source agentic video production system, with 100+ tools and 700+ skill files turning a coding agent into a full production studio.
thedotmack/claude-mem — +3,371 this week — compresses what an agent did during a session with AI and re-injects the relevant parts into its next one, now carrying a Vercel OSS Program badge.
mksglu/context-mode — +3,297 this week — sandboxes tool output to cut context usage and persist session memory across 17 agent platforms; its second appearance here, up from +2,241 a month ago.
anthropics/knowledge-work-plugins — +2,647 this week — Anthropic's own open-source release of 11 Claude Cowork plugins for specific roles — sales, legal, finance, product — each bundling the connectors and slash commands that role needs.
chthollyphile/folia-major — +2,504 this week — a local and Navidrome-compatible music player built around animated lyrics rendering, mostly documented and discussed in Chinese.
666ghj/MiroFish — +2,269 this week — a general-purpose swarm-intelligence engine for prediction tasks, backed by Chinese gaming company Shanda.
earthtojake/text-to-cad — +1,994 this week — gives a coding agent CAD-generation abilities via a Python package and skill, with its own Discord and a dedicated docs site.
remotion-dev/remotion — +1,803 this week — the established React-based framework for building videos out of code instead of a timeline editor, still climbing years after its first release.
vercel-labs/skills — +1,604 this week — Vercel Labs' CLI for the open agent-skills ecosystem, supporting OpenCode, Claude Code, Codex, Cursor and 75 more agents.