GitHub · AI Coding · AIGC

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

By 王若风 · Jul 16, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Skills are becoming a cross-platform protocol for agent behavior, and whoever sets that standard controls the next major developer ecosystem. Meanwhile, running a 744B MoE on a laptop removes the last cost barrier for using frontier open-source models against closed APIs.

Summary

GitHub's July 2026 trending data shows the Claude Code Skills ecosystem dominating the charts. Projects like mattpocock/Skills and obra/superpowers are pulling in tens of thousands of stars and forks, signaling that Skills are becoming a universal protocol layer across coding agents. The high fork counts indicate developers are not just starring these repos but actively modifying and deploying them.

Two other major threads emerged. JustVugg/colibri, a zero-dependency C engine, runs the 744B-parameter GLM-5.2 MoE model on a machine with just 25GB of RAM by streaming experts from disk. This decouples total parameters from active parameters, making consumer-grade inference for massive models a practical reality.

xAI's grok-build, a Rust-based coding agent harness, represents a shift in the agent arms race from model capability to the runtime environment. The simultaneous rise of safety guardrails and interaction-layer projects confirms that the infrastructure around coding agents is maturing rapidly.

Takeaways

— mattpocock/Skills gained 21,532 stars in one month, reaching 173,000 total stars, by open-sourcing a personal .claude directory of agent skills.

— obra/superpowers positions Skills as a full development methodology, not just a collection of prompts, and has 23,000 forks.

— Seven of the top ten monthly trending projects are directly related to Skills or coding agents.

— JustVugg/colibri runs the 744B-parameter GLM-5.2 MoE model in pure C with zero dependencies on 25GB of consumer RAM by streaming experts from disk.

— xai-org/grok-build, a Rust-based coding agent harness and TUI, gained 5,928 stars in its first week.

— Shell script has risen sharply in the language distribution because Skills are fundamentally organized as Shell and Markdown files.

— Rust now accounts for three entries each on the daily and weekly charts, driven by agent infrastructure projects.

— Sam Altman publicly endorsed open-source agent harnesses on the same day grok-build trended.

— Chinese developer content, including an open-source book on AI agents, is increasingly reaching global audiences through GitHub.

Conclusions

Skills are undergoing the same pattern as early mobile app stores: rapid quantity growth before a killer use case emerges. The real value will accrue to the framework that standardizes discovery and composition.

The decoupling of total parameters from active parameters in MoE models, exploited by colibri, changes the economics of model deployment. When a 744B model runs on a laptop, the cost argument for closed-source APIs collapses.

xAI and OpenAI both signaling interest in open-source agent harnesses suggests the next platform war will be fought over the runtime environment, not the model weights. The harness is the new browser.

The rise of Shell as an AI development language is a return to Unix philosophy: simple, composable text streams that any agent can execute. This undercuts complex SDKs as the default integration pattern.

Concepts & terms

Skills (in AI coding agents)

Reusable, packaged instructions—often written in Markdown or Shell—that extend a coding agent's capabilities for specific tasks, tools, or workflows. They are emerging as a cross-platform standard similar to MCP (Model Context Protocol).

MoE (Mixture of Experts)

A model architecture where only a subset of the total parameters (the 'experts') are activated for any given input. This allows models with hundreds of billions of total parameters to run with much lower computational cost by keeping inactive experts on disk and loading them on demand.

Agent Harness

A runtime shell or framework that hosts and manages a coding agent's execution, providing the environment, tool integrations, and user interface. It is distinct from the agent's underlying model and is becoming a new competitive layer.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗