Report by The Next New Thing.
Scroll for the report ↓
Ask an AI for a diagram and you get generic rounded boxes. This is a Claude Code skill that draws 39 kinds of editorial-quality diagram as plain HTML and SVG — no Figma, no build step.
View on GitHub →Architecture, flowchart, sequence, state machine, Sankey, Wardley map, fishbone and more — each in minimal light, minimal dark and full-editorial.
Point it at your site and it pulls the palette and type, so the diagram looks like the rest of your pages in about 60 seconds.
Every output opens straight in a browser. No JavaScript, no build step and no external image files.
cathrynlavery/openclaw-opsOpenClaw operations skill with health checks, repair scripts, watchdogs, update triage, and security scans.
cathrynlavery/codex-skillGive Claude Code a second opinion using OpenAI Codex - automatic plan review via hooks
cathrynlavery/codex-buildOrchestrator drives, Codex codes — execute an approved plan one task at a time with a test gate before every commit and exactly one PR at the end
A database that stores an AI agent’s memories, files and skills as folders under a viking:// path, so the agent browses its own context with ls and find instead of querying a black box.
Memories, resources and skills each get a viking:// URI, so an agent locates context the way a developer works with files.
Every entry is written as L0 abstract, L1 overview and L2 details, then loaded only as deep as the task actually needs.
Each query keeps its directory-browsing trajectory, so when a result looks wrong you can see exactly which path produced it.
volcengine/MineContextMineContext is your proactive context-aware AI partner（Context-Engineering+ChatGPT Pulse）
volcengine/ai-app-lab
volcengine/SearchCLIOpen CLI for integrating AI search, recommendation, and conversational retrieval into agent systems and business systems
A ready-made Arch Linux setup from David Heinemeier Hansson. One command turns a bare install into a fully themed, tiling, keyboard-driven desktop with every app and config already chosen for you.
View on GitHub →Window manager, terminal, editor, theme and hotkeys are all decided for you — the way Rails picks defaults so you do not have to.
The manual/ folder is its authoritative source and mirrors to learn.omacom.io, covering everything from Mac migration to dual boot.
DHH says 4.1 will bundle Theo’s T3 Code: “beefs are dumb, and we should all just be excited about the tech.”
Imagine giving your AI helper a big key ring. One connection lets it use thousands of apps, such as Gmail and Notion, to do everyday chores for you.
zapier.com/mcpIf you write software, this is the same key ring handed to your program instead of your chatbot — a few lines of code and your app can use those thousands of apps too.
zapier.com/sdkA 45-million-parameter model that only does tool calling, device control and structured extraction. The entire thing is one 14MB binary that runs a full session in about 28MB of RAM.
View on GitHub →Weights are baked into a single engine — no separate model files to manage, and inference makes no network calls at all.
A learned head returns a calibrated score, so you set a threshold, act above it and escalate below it.
A 256-token sliding window pins your tools as KV sinks, so total memory stays near 28MB however long the conversation runs.
cactus-compute/cactusQuantization, kernels, runtime and inference engine for mobiles, wearables, smart home and robots.
cactus-compute/cactus-hybridOn-device models that know when they're wrong: every answer carries a confidence score for cloud handoff.
cactus-compute/cactus-react-nativeCactus React Native package: Run AI locally in your React Native apps
Software that turns messy company data into a knowledge graph an AI agent can reason over — and records where every conclusion came from, so an auditor can trace a decision back to its source.
View on GitHub →The README calls it “The Open Source Palantir for AI Agents” — self-hostable, MIT-licensed and aimed at regulated industries.
Provenance is built in rather than bolted on, so you can show which records and which rules produced a given answer.
Supports both RDF and property-graph storage on W3C standards, with graph analytics and causal reasoning over the result.
A proxy that sits between your coding agent and the models. The agent keeps speaking its usual API while Switchyard translates the request and serves it from vLLM, Ollama, NIM or anything OpenAI-compatible.
View on GitHub →Converts among OpenAI Chat, Anthropic Messages and OpenAI Responses, so the agent never knows the backend changed.
LLM-as-classifier, signal-driven stage routing, escalation (cheap model first) and random splits for A/B tests.
The README carries its own warning — “Experimental software. Not for production use.” — and expects breaking changes before v1.0.
NVIDIA-NeMo/SpeechA scalable generative AI framework built for researchers and developers working on Large Language Models, Multimodal, and Speech AI (Automatic Speech Recognition and Text-to-Speech)
NVIDIA-NeMo/GuardrailsNeMo Guardrails is an open-source toolkit for easily adding programmable guardrails to LLM-based conversational systems.
NVIDIA-NeMo/DataDesigner🎨 NeMo Data Designer: Generate high-quality synthetic data from scratch or from seed data.
Mojo is a language that reads like Python but compiles to the speed of C++ and targets any vendor’s GPU. On 18 August 2026 its last closed piece — the compiler itself — went open source under Apache 2.0.
View on GitHub →The standard library was opened earlier; on 18 August 2026 the entire language, compiler included, landed under Apache 2.0.
Version 1.0.0 arrived in the MAX 26.5 release on 11 August 2026, four years and 36 releases after the project began.
Qualcomm completed its all-stock acquisition on 29 July 2026; founder Chris Lattner is now its EVP of Advanced AI Software.
modular/stack-prA tool for working with stacked PRs on github.
modular/llm-inference-handbookEverything you need to know about LLM inference
modular/mojo-gpu-puzzlesLearn GPU Programming in Mojo🔥 by Solving Puzzles
Give it a subject or a keyword and it writes the script, finds matching stock footage, generates the voice, subtitles and background music, and renders a finished vertical video in one pass.
View on GitHub →It writes the copy, pulls the footage, generates voice and subtitles, adds music and composites the final HD short.
At 112,986 stars it is among the most-forked AI projects on GitHub, which is the clearest signal of how many people run their own copy.
Ships a web interface and an HTTP API, and runs on Windows, macOS and Linux with Python 3.11+.
harry0703/AudioNotes快速提取音视频内容，整理成一份结构化的markdown笔记
harry0703/MangoDiskSafety-first disk cleaner and disk space analyzer for macOS and Windows, with duplicate file finder, large file cleanup, app uninstaller, and startup manager
harry0703/FlashVoice本地优先的 macOS 与 Windows 语音输入工具，支持端侧 SenseVoice 识别、全局语音输入、文件转写及可选 AI 校对。Local-first voice dictation for macOS and Windows with on-device SenseVoice, typing into any app, file transcription, and optional AI proofreading.
One hand-curated list of free public APIs — weather, finance, sports, government data, animals — grouped by category, with notes on whether each needs a key, allows CORS and supports HTTPS.
View on GitHub →Curated by hand since March 2016, making it one of the most-starred repositories on all of GitHub.
Each row lists the auth type — API key, OAuth or none — plus HTTPS support and CORS policy, before you write any code.
The README opens with an APILayer promotion above the list itself. The directory is free; its most valuable placement is not.
A command-line tool that checks one email address against more than 120 sites — Twitter, Instagram, Imgur and the rest — and reports which ones have an account attached to it.
View on GitHub →Returns a per-site result saying whether an account exists, plus any partly-masked recovery email or phone number the site reveals.
Each check uses the site’s own password-reset page, which is why the README notes the address is not notified.
Almost two years without an update, so expect some of the 120 site modules to have drifted out of date.
megadose/toutatisToutatis is a tool that allows you to extract information from instagrams accounts such as e-mails, phone numbers and more
megadose/ignorantignorant allows you to check if a phone number is used on different sites like snapchat, instagram.
megadose/OnionSearchOnionSearch is a script that scrapes urls on different .onion search engines.
A file-based notes app — backlinks, graph view, daily notes — where a Claude Code-style agent edits the same plain markdown, synced through your own GitHub repo.
View on GitHub →I think Shockwave is the most interesting thing I’ve used lately: local markdown with backlinks, graph view and daily notes, plus an agent working on the same files and your own GitHub repo for free sync and full commit history. Its companion server runs cron jobs and connects to Telegram, so you can reply to a scheduled message and resume the conversation with full context from your phone.
Building something? Send it in and it runs on a future show.
A full-featured agentic harness built local-first with security in mind — native multi-tier memory, switchable personas and per-persona tool lists.
View on GitHub →I’ve been following your videos and wanted to share Lumina, a feature-rich desktop agent I’m building local-first with security in mind. It has a native multi-tier memory architecture, switchable personas and tool lists, and a lot more.
Building something? Send it in and it runs on a future show.
One CLI that runs image prompts in parallel — pass -n 10 or a YAML batch of prompts and providers — talking straight to Gemini and OpenAI with your own key.
I built imagine because generating lots of images through web UIs meant clicking and waiting one at a time. It lets me run -n 10 or send a YAML batch of prompts and providers in parallel, turning twenty variations from half an hour into about a minute. It talks directly to Gemini and OpenAI with your own key, so there’s no aggregator markup, and I’m building video generation next.
Building something? Send it in and it runs on a future show.
Turns the daily TLDR newsletters into a chaptered podcast. Parsing and scripting are deterministic; a local 82M-parameter text-to-speech model reads exactly what TLDR published.
View on GitHub →I built TLDR Radio to turn the daily TLDR newsletters into a chaptered podcast, then wouldn’t let an LLM touch the text — parsing and scripts are deterministic, and a local 82M-param TTS just reads what TLDR published. Claude Code wrote most of it; it’s been building episodes unattended every night since July on a ten-year-old box in my basement.
Building something? Send it in and it runs on a future show.
One engine with many doors for providers, transports and tools — two switchable loops, a relevance gate, and a small-model-first path that escalates only when verification fails.
View on GitHub →I built aura-code as a framework for people to build on, not just use as-is — one engine with many doors for providers, transports and tools. It has two switchable loops (coding vs a cheap personal-assistant mode), a relevance gate that cuts tool-token use by about half, and a small-model-first path that only escalates when verification fails.
Building something? Send it in and it runs on a future show.
Rewrites business writing into Barbara Minto’s Pyramid Principle, with a consent gate and eight tests — three of which expect the skill to do nothing at all.
View on GitHub →I run Millwright Labs, and this is the first of the Agent Skills I’d been keeping private. It came from fact-checking an Instagram post whose “always give exactly three reasons” rule turned out not to be Minto’s. After its first version restructured a draft I’d only asked it to inspect, I added a consent gate and eight tests — three expect it to do nothing. GPT-5 and Gemini hold the line; Haiku doesn’t.
Building something? Send it in and it runs on a future show.
A reusable skill for stress-testing a strategy, tool, offer or product idea against five expert perspectives before you commit to it.
View on GitHub →I built the 5 Persona Advisory Board as a free, reusable decision-making framework for founders, creators and builders. Before they commit to a strategy, tool, offer or product idea, they can run it through five expert perspectives to catch execution risks, positioning problems, objections and capability gaps, and identify the strongest next move.
Building something? Send it in and it runs on a future show.
When ChatGPT, Claude, Gemini or Grok caps you mid-task, Pivvi passes the entire conversation to the next free model with headroom — no login, no API key.
View on GitHub →I watched the token-problems episode and built Pivvi for the other side of that problem: when ChatGPT, Claude, Gemini or Grok caps you mid-task, it hands the whole conversation to the next free AI with headroom — no login, no API key. There’s also Pivvi World, a globe of free AI tools so you can route around limits.
Building something? Send it in and it runs on a future show.
New hot repos every week — don't miss the next drop.
It tells YouTube to show this to more builders like you.
Tell me which repo you'll try first — or what to cover next.
New hot repos every week. Subscribe so you never miss the next drop.
