36,000+ stars in one week. Here are the 5 repos.

Tsinghua's THU-MAIC lab open sourced the classroom behind its 2024 'From MOOC to MAIC' paper, piloted on more than 500 students. Type a topic or upload a PDF, PowerPoint, audio or video and it builds a full course: outline, narrated slides, quizzes, live simulations, project activities. Then it casts the room. An AI teacher lectures with voice, draws on a whiteboard and points a laser at things. An assistant fills gaps. An AI classmate asks the questions you'd be too shy to. v1.0.0 on Aug 27 added a Pro workbench where you chat with an agent that plans a multi-lesson curriculum and edits pages with 20 built-in skills. I typed 'how does a database index work' into the hosted demo with no login and it cast Dr. Arthur Code and Maya Index and started lecturing on full table scans.

git clone https://github.com/THU-MAIC/OpenMAIC
cd OpenMAIC && cp .env.example .env.local
pnpm install && pnpm dev

Use it for: teaching yourself anything from a textbook you already own · a working example of multi-agent orchestration (a LangGraph director graph driving 28 action types) · exporting lessons as .pptx or .html. Honest flags: Node 22.19+ and pnpm 10+, the Windows install breaks on a cp step (issue #18), a Sep 1 issue reports quizzes grading the right answer wrong (#1326), and the discussion flow is scripted enough that one user called it an MVP. You can skip the install at open.maic.chat.

A plugin that makes your coding agent behave like the crustiest senior engineer in the building. Before it writes anything it climbs a 7-rung ladder: does this need to exist, is it already in the codebase, does the standard library do it, does the platform do it, does an installed dependency do it, can it be one line, and only then the minimum that works. It reads the code first and only then gets lazy about the answer. Six things stay permanently off the chopping block: trust-boundary validation, data-loss error handling, security, accessibility, hardware calibration, and anything you asked for by name. Dietrich Gebert's launch numbers on the same model and tests: no skill 3,629 lines, ponytail 490. His agentic benchmark on Haiku 4.5 got 54% less code. JetBrains ran their own test and got 15% less. I ran it on Sonnet 5 across two tasks and saw no difference, because Sonnet already skipped the over-build. The gap is real on weaker or more verbose models.

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

Use it for: teams on Haiku or GPT-class models where every ticket turns into 400 lines · /ponytail-review and /ponytail-debt on a repo you inherited. Those are two separate prompts in Claude Code, the README is explicit about it. Works as a plugin in Codex, Copilot CLI, Pi and OpenCode, and in about 20 other agents through AGENTS.md. Best line from the HN thread: 'The repo is bigger than most of the code Ponytail would allow me to write.'

A desktop app from Palash Debnath, a solo builder in India, that does what ElevenLabs charges $22 a month for, on your own machine. Clone a voice from a 3 to 15 second clip. Design a new voice from a text description of age, accent and style. Dub a video: it transcribes, translates, keeps each speaker's voice and re-synthesizes. System-wide dictation on a hotkey. EPUB or PDF in, .m4b audiobook out. 16 TTS engines and 11 ASR engines sit behind one Model Catalogue, and the default engine, k2-fsa OmniVoice, claims 646 languages. No account, no API key, no usage meter. It was renamed from OmniVoice-Studio on Aug 14, and v0.5.1 on Aug 28 turned it into a local speech platform other apps can call over HTTP, WebSocket, CLI or MCP.

# macOS Apple Silicon DMG / Windows MSI / Linux AppImage
https://github.com/debpalash/VoiceStudio/releases/latest

# or Docker
docker run -d -p 127.0.0.1:3900:3900 -v omnivoice-data:/app/omnivoice_data --name voicestudio palashdeb/omnivoice-studio:stable

Use it for: dubbing your own reels into Spanish or Hindi in your own voice · audiobooks from PDFs you already own · dictation that never leaves your laptop. The catch that matters: the default OmniVoice weights are CC-BY-NC, so anything you sell needs one of the commercial-OK engines like CosyVoice 3, GPT-SoVITS or VoxCPM2. Apple Silicon only on Mac, Intel Macs can't run the local backend. 8GB RAM and 10GB disk minimum, and it's an active beta, so use tagged releases.

Matt Pocock (Total TypeScript, AI Hero) made his working .agents directory public: 37 skills he runs day to day with Claude Code and Codex. The flagship is grill-me, which interrogates every assumption in your plan before a line gets written. His own post says it's the most useful skill he's written and he uses it outside coding. to-spec turns a conversation into a spec, to-tickets breaks it into tracer-bullet tickets, implement builds and closes with code-review. wizard is the odd one: it hands you a disposable bash script to walk through the steps only a human can do, dashboards and secrets and CI, with a rule to never invent a step. resolving-merge-conflicts bans --abort outright. He positions the set against GSD, BMAD and Spec-Kit: small, composable, works with any model.

claude plugins install mattpocock-skills

# or an editable copy. Pick ONE, both installs = every skill twice
npx skills@latest add mattpocock/skills

Use it for: stealing grill-me and wizard even if you skip the rest · a reference for how a working engineer structures SKILL.md files. It leans toward his TypeScript and web stack, /code-review clashes with Claude Code's built-in command (issue #483), and the HN thread is split. One commenter, in full: 'Christ, this is exhausting.'

163 skills plus one lookup skill covering 78 public databases, PubMed, PubChem, ChEMBL, ClinicalTrials.gov, that turn Claude Code, Codex, Cursor or Gemini CLI into a research assistant. Each skill is a SKILL.md instruction manual telling the agent the documented way to call a database, with a built-in rule to never follow instructions embedded in returned data. The README's showpiece chains a whole drug-discovery pipeline in one prompt: pull EGFR inhibitors from ChEMBL, analyze with RDKit, generate analogs, dock them against an AlphaFold structure with DiffDock, search PubMed, write the report. I ran it with zero API keys and asked for the newest papers on CRISPR base editing off-target effects. It hit PubMed live, got 419 records, noticed its first query was too broad, re-ran it with title and abstract fields, and graded which results were on topic. The team put the paper behind it on arXiv on Aug 30 and linked it from the README on Sep 2.

npx skills add K-Dense-AI/scientific-agent-skills

Use it for: literature pulls with real provenance (endpoints, parameters, access date) · the scientific-writing skill, which refuses to invent citations. Flags: the '190,000 scientists' line is K-Dense's own unsourced number, the README funnels to their paid app, some databases need their own keys, and community skills are reviewed 'with limited resources'.