20,000+ stars in one week. Here are the 5 repos.

An Agent Skill for Claude Code, Codex, Factory Droid, and Pi that generates 39 diagram types as self-contained HTML and SVG. It reads your website first, pulls your palette and fonts into semantic tokens, runs contrast checks, and then every diagram you ask for comes out matching your brand. It can also redraw existing draw.io or Mermaid diagrams into the same system, and named client profiles let one install serve multiple brands.

/plugin marketplace add cathrynlavery/diagram-design
/plugin install diagram-design@diagram-design

Use it for: architecture and flowchart diagrams for posts, decks, and docs without opening Figma · redrawing old Mermaid slop into something you'd actually publish · agencies running multiple client palettes from one install. This is the one I'd install today.

DHH's opinionated take on Linux: Arch plus Hyprland, shipped as a bootable ISO with full-disk encryption on by default. Version 4.0 rewrote the whole desktop shell into one Quickshell process and treats AI agents as system citizens, down to a Claude Code weekly-limit widget in the bar. A week after it shipped, the Omacom Foundation launched with $8M from 8 founding patrons at $1M each: DHH, Tobi Lütke, Patrick Collison, Michael Dell, Jack Dorsey, Matthew Prince, Brendan Iribe, and Jason Fried.

https://iso.omarchy.org/omarchy-4.0.0.iso

Use it for: turning a spare laptop or a Framework into a real dev machine · leaving macOS with a setup that works out of the box · tiling-window-manager life without weeks of dotfile tuning. Heads up: the full-disk install wipes the target drive, and you need Secure Boot off.

Needle 2 is a 45 million parameter tool-calling model quantized to about 2 bits per weight, so the whole thing is a 14MB file running in roughly 28MB of RAM. It skips chat entirely. You hand it tools with a Python decorator and it maps a spoken or typed sentence to a typed function call, on device, with zero network calls at inference. When it's unsure, it escalates to the cloud.

pip install cactus-needle

Use it for: voice control in smart-home gear, wearables, and robots that have to work offline · sub-$200 Android phones where every MB counts · Pi and microcontroller projects, ESP32-S3 class and up. Fine-tune it on your own device functions, the base model is deliberately narrow.

A context database from ByteDance's cloud arm that stores agent memory, RAG documents, and skills as one virtual filesystem. Your agent browses viking:// paths with ls, tree, find, and grep instead of querying a vector-store black box, and every entry loads in three tiers: a 100-token abstract, a 2k overview, then full content only on demand. Sessions automatically distill into long-term memory after they close. Their own benchmark took Claude Code from 57% to 80% on long-conversation recall while cutting input tokens by up to 91%.

pip install openviking --upgrade

Use it for: agents that forget everything between sessions · token-heavy RAG where full documents keep landing in context · retrieval you can actually debug, since it leaves a readable browsing trail. Two flags before production: the benchmark numbers are ByteDance's own, and the license is AGPL-3.0.

A Rust proxy from NVIDIA that routes LLM calls across models and providers while your app keeps speaking plain OpenAI or Anthropic API format. Its escalation router starts every agent turn on a cheap model and hands off to a frontier one only when a judge flags sustained difficulty. In NVIDIA's LangChain eval across 145 agent tasks, that cut costs 74% with only 7% of calls ever touching the big model.

cargo install --locked switchyard-server

Use it for: mixed model fleets where frontier spend hurts · A/B testing models behind one endpoint · keeping native Anthropic and OpenAI formats with zero client rewrites. It's pre-alpha and the API will change before v1.0, so I'd wait unless you enjoy the bleeding edge. LiteLLM and OpenRouter are already building it in as a plugin.