Selection criteria: Trendshift daily star gain (stars today) is the primary metric. Repos whose recent growth has slowed or flattened are excluded no matter how large the cumulative count. Repos featured yesterday return only if their gain held steady or widened.
xai-org/grok-build — +9,500 ⭐ today (13,000 total)
https://github.com/xai-org/grok-build
This is xAI (SpaceXAI) releasing the full source of Grok Build — the terminal AI coding agent behind its grok CLI — under Apache 2.0. It is roughly 844,530 lines of Rust containing the entire harness that wraps the model: the agent loop that assembles context, calls the model, parses the reply and dispatches tool calls; the tool implementations for reading, editing and searching code and running shell commands; and a full-screen TUI with scrollback, prompt, modals and an inline diff viewer. The crate layout splits cleanly by role, which makes it unusually readable as a reference: xai-grok-pager is the TUI, xai-grok-shell is the agent runtime plus the leader/stdio/headless entry points, xai-grok-tools holds the tool implementations, and xai-grok-workspace handles host filesystem, VCS, execution and checkpoints. Beyond the interactive TUI it runs headless for scripting and CI, or embeds in editors via the Agent Client Protocol (ACP), and the extension system for skills, plugins, hooks, MCP servers and subagents ships with it. The surge traces back to security researchers revealing that Grok Build had been uploading developers’ entire Git repositories — commit histories and credentials such as API keys and SSH keys included — with the source landing right after Elon Musk announced all uploaded user data would be deleted, drawing a wave of verification traffic. Worth noting: external contributions are not accepted (per CONTRIBUTING.md), and the model itself stays in xAI’s cloud, making this an open-source client.
Practical angle: you can read a production coding agent’s harness design — context assembly, tool dispatch, checkpointing — as a complete reference implementation, which makes it a useful benchmark when designing an in-house agent runtime.
Tags: #AgentHarness #RustLang #TerminalUI #CodingAgent #MCP #AgentClientProtocol
Fei-Away/Codex-Dream-Skin — +5,200 ⭐ today (5,200 total)
https://github.com/Fei-Away/Codex-Dream-Skin
A theming tool that swaps out the entire UI of the Codex desktop client. The injection method is the whole point: it never touches the official .app, app.asar or the code signature, instead injecting custom UI through local loopback CDP (Chrome DevTools Protocol). Hand it any image and it turns that image into one continuous full-window theme running beneath the native sidebar and main surface, with route-aware translucency keeping the home, task, plugin, scheduled-task and pull-request controls readable and fully interactive. A fresh install seeds the theme library with one tested featured preset plus five procedural abstract presets generated with pure Node + zlib — no photos, no third-party art or likeness. It ships one-click install and restore scripts for macOS and Windows separately, and if you simply hand the repo address to Codex it will clone the project, run the install scripts and launch the skinned Codex for you. The spike came from Chinese resale site Xianyu, where Codex themes were being sold for 66–99 yuan; once that hit X, the reaction that a free open-source alternative already existed drove a flood of traffic in a single day.
Practical angle: the CDP injection pattern itself — theming without touching signatures or packages — is a reference for applying brand themes to internal Electron tools without breaking integrity checks.
Tags: #CDPInjection #ElectronTheming #DesktopSkinning #CodexClient
Nutlope/hallmark — +2,400 ⭐ today (9,300 total) · 3 days running
https://github.com/Nutlope/hallmark
An anti-AI-slop design skill that forces the UIs generated by Claude Code, Cursor and Codex to stop looking machine-made. It works as a scoring and revision loop: output is graded 1–5 on six axes — Philosophy, Hierarchy, Execution, Specificity, Restraint and Variety — and anything below 3 on any axis triggers another revision pass. On top of that sit 57 slop-test gates that block the model from falling back to the defaults every LLM absorbed during training. The differentiator is that it demands structural variety rather than the visual kind: two pages from two different briefs must not share the same hero → 3-feature → CTA → footer rhythm, and should feel like different sites rather than colour swaps of one template. It ships twenty themes and four verbs, cyclable with the T key, and when a brief carries creative intent that no catalog theme fits, it switches to Custom and designs the page from scratch with a made-to-measure palette, type and layout — with no template underneath but the same 57 gates still applied. After first appearing at #8 on Trendshift’s daily list on July 13 it climbed to #2 on GitHub Trending on the 15th, and its daily gain has widened three days straight: +853 (Jul 15) → +967 (Jul 16) → +2,400 (Jul 17). MIT licensed.
Practical angle: it catches the problem where every agent-generated landing page and dashboard looks alike, via a scoring loop — easy to bolt on for teams drafting customer-facing UI with AI.
Tags: #DesignSkill #AntiSlopDesign #UIGeneration #AgentSkills
bojieli/ai-agent-book — +659 ⭐ today (1,300 total)
https://github.com/bojieli/ai-agent-book
The open-source main repository for Understanding AI Agents in Depth: Design Principles and Engineering Practice (《深入理解 AI Agent：设计原理与工程实践》) by Bojie Li (李博杰). The point is that the book’s full text, the compiled PDF and the per-chapter companion code are all published outright: the book/ directory holds the source for introduction.md, chapter1.md through chapter10.md and afterword.md, alongside the finished 深入理解-AI-Agent-李博杰-v1.0.pdf. The whole book is built around a single formula — Agent = LLM + Context + Tools — and ten chapters carry it from agent fundamentals through to the concept of harness engineering. What separates it from the usual “intro to agents” curated link list is that every chapter has runnable code attached, so you can execute as you read rather than collect bookmarks. Of its 1,300 cumulative stars, 659 landed today — effectively a repo discovered this morning.
Practical angle: usable as-is as a free textbook for an internal agent study group or new-hire onboarding curriculum, with theory and runnable code paired chapter by chapter (note: the original text is in Chinese).
Tags: #AgentEngineering #HarnessEngineering #ContextEngineering #TechnicalBook
tandpfun/wardrobe — +569 ⭐ today (569 total)
https://github.com/tandpfun/wardrobe
A tool that pulls the clothes out of your photos automatically, organizes them into a personal wardrobe catalog, and then generates outfits from it. It works in two stages: the OpenAI Responses API detects every garment in a photo, and the OpenAI Images API (gpt-image) extracts a clean product cutout of each item. The repo ships two Codex skills — $import-clothes, which imports clothes, generates modeled photos and adds them to the wardrobe, and $generate-outfits, which styles complete outfits from the wardrobe and generates a modeled lookbook. Setup is clone the repo, npm install, configure .env, then npm run dev and open localhost:5173; the importer stays disabled until you add OPENAI_API_KEY to .env and drop a PNG reference photo at data/model-reference.png. The creator’s demo — giving GPT-5.6 Sol access to his camera roll, having it extract every piece of clothing he owns from his photos, then rendering new outfits onto himself — spread on X, and after being open-sourced yesterday all 569 of its cumulative stars arrived today.
Practical angle: the pipeline — photo → individual object detection → background-free catalog image → combination generation — ports directly to commerce product onboarding and inventory cataloging.
Tags: #ImageSegmentation #GPTImage #ResponsesAPI #CodexSkills
Today’s Summary
xAI’s Grok Build took a commanding first place at +9,500 stars in a day, and the rise looks driven largely by verification traffic in the wake of the unauthorized repository-upload controversy. Codex theming tool Codex-Dream-Skin stacked its entire cumulative count (+5,200) in a single day as the paid-skin backlash worked in its favour, while anti-AI-slop design skill hallmark widened its daily gain for the third consecutive day — +853 → +967 → +2,400 — making it the only sustained climber on the board.
Daily AI Brief 
Leave a comment