Transcription
The monthly GitHub roundup is here. After yesterday's video, a bunch of you said, "Why not do this monthly?" Well, say less. Today kicks off GitHub trending monthly number one. 20 open source projects, all bangers, all from the past month. If you're building, coding, or shipping, you're going to love this list. Grab your caffeine. Let's jump straight into the best repos of the month.
What if you could build your own chat GPT for a hundred bucks? Yeah, this is Nanohat by Andre Karpathy, a mini chat GPT pipeline in just 8,000 lines of code. It already has 35,000 stars on GitHub and it's blowing up. You get everything. Tokenizer, pre-training, fine-tuning, reinforcement learning, web UI, literally the whole stack. Fire up a GPU, run one script, and after just four hours, boom, you can chat with your own private AI.
It trains on legit data sets like fine web edu, small talk, gsm8k, and it even runs code safely in a sandbox. This is deepseek OCR and it's way bigger than just OCR. 19,000 stars on GitHub and it's already rewriting the multimodal playbook. It takes images, invoices, research papers, handwritten notes, and converts them into clean structured text with crazy accuracy. But here's the wild part. Instead of bloated vision tokens, it compresses them 10 times smaller than text tokens. Meaning you can store more knowledge in vision form than text. Upload entire books, whole code bases, company docs, and your AI remembers all of it.
What if JSON didn't burn tokens like crazy and you could feed way more structured data to your LLM for the same price? Yeah, like literally cutting your prompt cost in half. This is tune token oriented object notation. Think JSON for LLMs, but way more compact and it already has devs buzzing. Instead of repeating keys over and over again like JSON, Tune lets you define them once, then stream rows of data. Super clean, super tight. It borrows YAML style indentation, mixes in CSV style tables and boom, 30 to 60% fewer tokens in your prompts. So instead of this chunky JSON block, you get something tiny, readable, and LLM friendly.
Okay, so what happens if you let five AIs battle it out on Wall Street? No humans allowed. That's AI trader, and it's wild. This project has 8,200 stars on GitHub and throws AI agents into the NASDAQ 100. Each one gets $10,000, real market data, and one mission. Beat the market or go broke trying. They read financial news, analyze stocks, execute trades, all without touching a single human brain cell. No prompts, no programming during the match. Just pure autonomous strategy. GPT Claude, Quen, whoever you bring, they can join the arena. Charts, logs, profit graphs, every move tracked, every trade recorded.
Everyone talks about AI replacing developers, but this project, it upgrades developers instead. It's called Superpowers. 5600 stars on GitHub, and it's basically a skill tree for AI coding assistants. Not prompts, not snippets, actual engineering superpowers. Think test-driven development on autopilot, async testing patterns, systematic debugging, root cause tracing, code review workflows, even parallel agent collaboration. Yeah, your AI can spin up sub agents like it's managing a dev team. It loads proven workflows automatically and forces consistency. No guessing, no ad hoc coding, pure engineering discipline.
What if you could have studio quality voices right on your phone? This is New TS Air by Newonic. 3800 stars on GitHub. The super realistic on-device TTS model with instant voice cloning built on a 0.5B LLM backbone. It runs on phones, laptops, even Raspberry Pies. Fast, light, and secure. You can clone a voice in just 3 seconds of audio, and it sounds shockingly human. It uses new codec, a neural audio codec that crams amazing quality into tiny files. And with a 248 token context window, it can handle 30 seconds of audio in real time.
Meet Skillseker. 3100 stars on GitHub. The tool that scrapes, analyzes, and packages everything for Claude AI automatically. It reads docs, parses code deeply, detects conflicts, organizes topics, and even extracts examples, all into a ready-to-use skill. You can feed it game engine docs, frameworks like React or Django, internal company manuals or PDFs. It handles all sources. Combine sources, spot outdated docs, fill gaps, and create one single source of truth. No manual labor required. It even boosts your skills with AI enhancement. All local, no API costs.
What if you could edit, merge, and secure PDFs without ever uploading a file to the cloud? Yeah, total privacy right in your browser. This is Bento PDF. 2700 stars on GitHub. A client-side privacy-first PDF toolkit. Merge, split, reorder, crop, rotate, posterize, even convert images or markdown to PDF. All locally. No servers, no limits. Want to fill forms, flatten annotations, redact sensitive info, or add signatures? Done. Need OCR for scanned documents, compress huge files, or change metadata? All there. Built with modern web tech and handles even massive PDFs effortlessly.
Ever wished your financial research assistant actually thought like an analyst instead of just giving cute chatbot answers? Meet Dexter, an autonomous finance research agent designed to think, plan, and self-correct like a Wall Street pro. This isn't GPT, but for stocks. It breaks complex financial questions into steps, pulls live market data, verifies its own analysis, fixes mistakes, and only stops when it has a confident data-backed result. Imagine asking, "Is Nvidia still undervalued after earnings?" Dexter builds a plan, fetches the numbers, runs the logic, checks itself, and delivers a real thesis, not hype.
Da Surf, your AI-powered research workspace that pulls files, PDFs, YouTube videos, and the entire web straight into your stream of thought. No more juggling tabs, screens, copy-paste chaos. Surf turns your computer into a thinking machine. It runs on Mac, Windows, Linux, and keeps everything local. Ask a question while watching a YouTube video. Boom. Timestamped notes. Open a PDF. Get structured answers with page citations. Say, "Search this." Surf does research in the background and injects results into your notes. And the wild part, you can literally tell it to generate mini apps inside your notes. Surflets. No coding required.
Beads gives AI coding agents a memory. Not just a scratch pad, but a full graph-based issue system that chains tasks together like well, beads. Instead of dumping half-baked plans in Markdown, your agent files proper epics, tasks, dependencies, and even logs discoveries as they go. That means no more forgotten bugs, derailed road maps, or "where were we again?" moments after compaction. You don't use it directly. Your agent does. Just install beads. Tell your agent to use the BD tool and suddenly they can manage long-term multi-session work across machines.
This repo hosts Pokey's 7 billion parameter deep research agent designed to answer complex questions using real-time web search and content extraction, not hallucinations. Pokei doesn't just fetch links. It performs multi-turn research, searching, reading, summarizing, and iterating until it builds a verified answer. It's tested on benchmarks like HLE, Gaia, and Browse Comp and delivers state-of-the-art results in deep web reasoning. Under the hood, it integrates Serper for search, GINA for content extraction, and Gemini for summarization and evals, all wrapped in a scalable 7B architecture that can run on a single A100 GPU.
Oxdraw, a code tool that turns mermaid syntax into a fully editable visual canvas and syncs every change back into your source file. Think of it as VS Code for diagrams, declarative, versionable, reproducible, but with real drag-and-drop precision. No more abandoning Mermaid because the arrows look janky. Ox gives you precise node control, custom connectors, color pickers, manual edge routing, even grid snap, drag, and alignment guides. Every tweak you make visually persisted as comments in your MMD file. So the diagram stays deterministic, readable, and git friendly.
Imagine a trading system where AI agents don't just trade, they compete, evolve, and level up like game characters. That's Nox, a universal agentic trading OS that just closed the loop in live crypto markets. And now it's expanding to stocks, futures, and forex. This isn't chatbot trades BTC. This is multi-agent self-play, low latency execution, unified backtesting and risk controls all in one architecture. Agents battle each other in real time, learn from losses, reinforce winners, and adapt strategy based on live P&L. Binance, Hyperliquid, Asterex, plugins, switch markets, keep the same AI brain.
What if you could turn a 20-page research paper into a full presentation video without touching PowerPoint, Keynote, or your mic? Meet paper-to-video. The first data set and system that actually auto-generates academic videos from papers. It reads the PDF, builds slides in latex, fixes layouts, writes subtitles, moves the cursor, clones your voice, and even renders your talking head. Think, "Hey AI, present my Nurip." And it's done. Meanwhile, a benchmark judges how well the video explains the science, not just how pretty it looks.
Open Memory. Long-term semantic memory for any AI system. Preferences, past conversations, code you've touched, files you've edited, all recalled instantly. Drop it in and boom, your AI finally stops forgetting who you are. This isn't a vector DB hack. It uses a hierarchical biologically inspired memory structure. Episodic, semantic, procedural, even reflective memory nodes. One canonical memory per idea, no duplication, fast recall, and explainable reasoning.
[Music] Ever wished you could just point at a button in your app and say, "Hey, Claude, fix this." Yeah. No more explaining DOM paths, no copying JSX, no guessing. This is React Grab, a tiny script that turns your UI into something your AI can literally grab with a click. Hold command C, click any element in your app, and boom, cursor or claw code instantly gets the exact component. Real code, real context. No spelling through files. It's like giving your coding agent X-ray vision for your UI. You click, it knows. You say, "Change this navbar." It edits the right component. First try.
Networking Toolbox. The offline-first networking Swiss Army knife with 100+ tools for diagnostics, conversions, configs, and sanity checks. Ping, DNS lookup, subnet calculators, SSL checks, encoders, decoders. Literally everything you'd normally hunt across random CLI tools and janky websites, all running locally, even on a plane, in a data center, or in a panic at 3 a.m. when the internet dies. Customize layouts, themes, bookmarks. Make it your own command center. It has 1300+ stars on GitHub.
Ever wished you could monitor all your NVIDIA GPUs without SSH, terminal chaos, or clunky software? Meet GPU hot, a real-time web-based GPU dashboard that shows every metric you care about instantly. Utilization, temperature, memory, power, clocks, fan speed, everything in seconds. One server, dozens of GPUs, hundreds, even scale without breaking a sweat. See which processes are hogging memory, track historical charts, monitor throttling, all from your browser. No setup, no config, no SSH nightmares. Just one command and your GPU fleet is live.
Ever wanted to build your own AI agent workflows without writing a single line of glue code? Meet Open Agent Builder, a drag-and-drop visual workflow builder for AI pipelines powered by Firecrawl. Design complex multi-step agent workflows in minutes and watch them execute live with streaming updates. Perfect for web scraping, automated research, content generation, or even data transformation. You just drag nodes, connect them, and go. Start, agent, transform, if else, loops, human approval, every node you need.