跪拜 Guibai
← Back to the summary

Skills Are Eating the GitHub Trending Charts, and a 744B MoE Just Landed on a Laptop

I pulled a batch of GitHub trending data across three dimensions: daily, weekly, and monthly. In July 2026, the open-source landscape had a very clear main thread: the Skills ecosystem is exploding, large-model consumer-grade inference is landing, and the Coding Agent harness is becoming a new battleground.

One-Sentence Conclusion

The Skills ecosystem spawned by Claude Code is taking over the GitHub trending charts. An entire category is exploding simultaneously: mattpocock/Skills, obra/superpowers, Graphify-Labs/graphify, Nutlope/hallmark, coreyhaines31/marketingskills. These five names appear repeatedly across the daily, weekly, and monthly charts, with keywords all being claude, agent, skill, code.

If I had to give this month a theme word, it would be Skills.

Monthly Chart: The Full Takeover of the Skills Ecosystem

Let's look at the monthly chart first (June 16 ~ July 16), as it best illustrates structural trends.

Top 1 is mattpocock/Skills, which gained 21,532 stars in a month, reaching a total of 173,000 stars. The description is simple: "Skills for Real Engineers. Straight from my .claude directory." Basically, he open-sourced the Skills from his personal .claude directory. This alone gained over 20,000 stars in a month. The signal is clear: the developer community has a massive information hunger for "how to write and organize Skills."

Even more noteworthy is the Top 4 entry, obra/superpowers (+13,675, total 257,000 stars), an "agentic skills framework & software development methodology." Note its positioning: not a single Skill, but a complete development methodology framework. This means Skills are evolving from "scattered prompt fragments" into "systematic engineering practices."

Among the monthly Top 10, 7 projects are directly related to Skills/Agents:

Rank Repository Monthly Star Gain Positioning
1 mattpocock/Skills +21,532 Collection of engineer Skills
2 MadsLorentzen/ai-job-search +19,341 Job search framework based on Claude Code
4 obra/superpowers +13,675 Agent Skills development methodology
5 Graphify-Labs/graphify +13,405 Code knowledge graph Skill
6 JuliusBrussee/caveman +11,983 "Caveman" Skill saving 65% tokens
7 langchain-ai/openwiki +11,728 Agent auto-documentation CLI
10 emilkowalski/Skills +9,714 Design/Engineering Skills

This wave is not a single project's accidental hit; it's a collective explosion of an entire category. Moreover, the fork counts for these projects are absurdly high: mattpocock/Skills has 15,000 forks, obra/superpowers has 23,000 forks, and msitarzewski/agency-agents has 21,000 forks. High fork counts mean people aren't just "looking"—they are genuinely taking them, modifying them, using them, and building on them.

Why Skills, and Why Now

My previous judgment is being validated by data: Skills are becoming a universal protocol layer across Agents. Look at the description for Graphify-Labs/graphify: "AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more)." One Skill is simultaneously compatible with five or six mainstream Agents. The Skills standard is breaking away from a single platform and becoming an open protocol, much like MCP.

From the trend data, Claude Code's Skills standard currently leads, but other platforms (Codex, Cursor, Gemini CLI) are all adapting to it. Whoever becomes the de facto standard first controls the entry point for the next-generation Agent ecosystem. This layer is more worth watching than the models themselves.

Weekly Chart: Two New Main Threads Emerge

On the weekly chart (July 9 ~ July 16), the Skills ecosystem remains strong, but two new main threads have appeared.

Thread One: Consumer-Grade Large Model Inference—JustVugg/colibri

Weekly Top 1 is JustVugg/colibri, gaining 8,051 stars in a week, running GLM-5.2 (744B MoE) on a consumer-grade machine with 25GB of memory using pure C.

A few key points:

The key technical point is: the MoE architecture decouples "total parameters" from "active parameters." For a 744B model, a single inference might only activate a few tens of billions of parameters; the remaining expert parameters sit on disk and are loaded as needed. This makes running ultra-large models on consumer hardware possible.

This project appears on both the weekly and monthly charts (monthly +14,428), so it's not a flash in the pan. Coupled with the recent strength of open-source large models like GLM, Deepseek, and Kimi, consumer-grade inference engines are becoming a clear technical direction.

Thread Two: Coding Agent Harness—xai-org/grok-build

Weekly Top 4 is xai-org/grok-build, gaining 5,928 stars in a week. It's a new project, so its total stars are exactly 5,928, all gained in one week.

This is an official release from xAI (Elon Musk's AI company), positioned as a "coding agent harness and TUI" (full-screen, mouse-interactive, extensible). Written in Rust, performance and reliability are priorities.

The word "Harness" is key. It's not another Coding Agent, but a shell/framework that hosts the Agent's operation, akin to the IDE kernel of the terminal era. xAI entering this space indicates that the competition for Coding Agents has extended from "model capability" to "runtime environment."

Interestingly, Sam Altman said on X the same day that there is "a reason to favor open-source harnesses," publicly supporting open-source harnesses. The CEO of OpenAI supporting something xAI is doing? Consider the temperature of the water in this race.

Daily Chart: New Faces and Sustained Heat

The daily chart (July 16) best reflects "what is happening right now."

Daily Top 1 is xai-org/grok-build (+3,814), continuing its triumphant advance. A new project gaining thousands of stars daily for multiple consecutive days is a top-tier performance curve in the open-source world.

A few new faces on the daily chart worth naming:

Commonality Across Three Charts: Who Can Survive the Cycle

Overlaying the daily, weekly, and monthly charts to see which projects appear on all three reveals the truly structural trends, not just short-term traffic.

Projects appearing on all three charts:

Project Daily Weekly Monthly
mattpocock/Skills ✓ (Top 1)
xai-org/grok-build ✓ (Top 1)
Graphify-Labs/graphify
Dicklesworthstone/destructive_command_guard
JustVugg/colibri ✓ (Top 1)

mattpocock/Skills and Graphify-Labs/graphify are the only two Skills-type projects appearing on all three charts, indicating their attention is the most sustained. The former is the benchmark for "Skills collections," and the latter represents "Skill as a tool" (turning a codebase into a knowledge graph).

xai-org/grok-build and JustVugg/colibri didn't make all three charts, but each took Top 1 on the charts they did appear on: grok-build was the daily champion, and colibri was the weekly champion. These two represent the two new main threads outside of Skills: the Coding Agent harness and consumer-grade inference.

Language Distribution: The Unexpected Rise of Shell

Looking at the combined language distribution across the three charts reveals a counter-intuitive finding.

Python and TypeScript remain the mainstays, which is expected. But the proportion of Shell has risen significantly: 2 on the daily chart, 2 on the weekly chart, and 3 on the monthly chart.

Why? Because Skills are fundamentally organized using Shell/Markdown. mattpocock/Skills is Shell, obra/superpowers is Shell, and msitarzewski/agency-agents is also Shell. The explosion of the Skills ecosystem directly elevated Shell's weight on the trending charts.

This is an interesting signal: in 2026, writing Shell scripts is no longer just "something ops people do"; it has become one of the mainstream ways to build AI Agent capabilities. Skills organized in Shell can naturally run in any environment with bash installed, a fundamental advantage for cross-platform compatibility.

Rust's performance is also worth noting: 3 on the daily chart, 3 on the weekly chart. grok-build (Rust), destructive_command_guard (Rust), openinterpreter (Rust). The infrastructure layer for Coding Agents is migrating towards Rust, pursuing performance and security.

The Presence of Chinese Developers

Monthly Top 12 is an old acquaintance, 1c7/chinese-independent-developer (+7,094, total 56,000 stars), a list of Chinese independent developer projects. This project's long-term presence on the trending charts shows that overseas interest in the Chinese independent developer ecosystem is indeed growing.

There are also a few Chinese-language projects on the daily chart: bojieli/ai-agent-book (Li Bojie's open-source book "Understanding AI Agents in Depth"), dimthink/PriceAI (an AI subscription card price comparison tool), Yu9191/wloc (Apple network location spoofing). Chinese technical content on the AI Agent theme is reaching overseas readers via GitHub.

My Take

Frankly, after reviewing this data, I have three judgments.

First, Skills are not a gimmick; a real paradigm shift is happening. Over 20,000 stars and tens of thousands of forks in a month cannot be generated by hype alone. Developers are genuinely using Skills to reorganize their workflows. This category is still in its early stages; current Skills are like the App Store in 2010—the quantity is exploding, but a true "killer Skill" has yet to emerge.

Second, the competition for Coding Agents has entered the "deep water zone." As model capabilities converge, the competitive focus shifts to the harness (runtime environment), guard (safety guardrails), and skin (interaction layer). grok-build, destructive_command_guard, and Codex-Dream-Skin appearing on the daily chart on the same day is not a coincidence; it's the collective maturation of the Agent infrastructure layer.

Third, open-source large models + consumer-grade inference is the major thread for the next six months. colibri proved that a 744B MoE can run on a laptop. When the cost of running drops to zero and the barrier to entry drops to zero, the impact of open-source models on closed-source APIs will truly begin.

Today's Keywords: Skills · Harness · Consumer-Grade Inference · MoE Streaming


Data source: GitHub Trending + Trendshift, captured at 2026-07-16 14:50 UTC+8. See the three reports in the github-trending-archive/ directory for the full charts.