GitHub · Developer · Agent

Agent Skills Are Eating GitHub: The July 2026 Open-Source Landscape

By 怪侠说不说 · Jul 26, 2026

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

The standardization of Agent Skills and MCP infrastructure means the competitive moat is shifting from which model you use to how you configure and orchestrate it. Developers who treat Skills as composable, shareable engineering artifacts will outpace those still prompting by hand, much as infrastructure-as-code displaced manual server configuration.

Summary

July 2026 marks a paradigm shift on GitHub where Agent Skills — structured prompt templates, tool configurations, and output constraints — have become the dominant open-source category. Three major repositories from Matt Pocock, obra, and Addy Osmani represent competing philosophies for standardizing how developers interact with coding agents, collectively amassing hundreds of thousands of stars and forks.

The MCP (Model Context Protocol) infrastructure layer is also maturing rapidly, with projects like codebase-memory-mcp building persistent knowledge graphs for codebases and Agent-Reach giving agents eyes on the open web. This infrastructure build-out, spanning perception, memory, orchestration, security, and communication layers, signals that agents are moving from chat interfaces to production-grade workhorses.

Beyond coding, AI tools are industrializing creative work. OpenMontage provides an open-source agent-driven video production system with 12 pipelines, while OpenCut challenges CapCut's subscription model. In finance, Vibe-Trading and ai-hedge-fund lower the barrier to quantitative trading through natural language interfaces and multi-agent collaboration, respectively.

Takeaways

— Three Agent Skills repositories — mattpocock/skills, obra/superpowers, and addyosmani/agent-skills — collectively gained nearly 500,000 stars on July 11, 2026.

— Skills are structured combinations of prompt templates, tool-calling configurations, and output format constraints, analogous to Dockerfiles for agent behavior.

— mattpocock/skills provides 21 small, composable Claude Code skills including grill-me, which forces the agent to interrogate the developer before coding.

— obra/superpowers enforces a strict Spec → Plan → Implement → Review → Ship methodology rather than offering pick-and-choose skills.

— addyosmani/agent-skills bakes in production-grade quality gates: OWASP Top 10 checks, Lighthouse performance budgets, and WCAG 2.2 AA accessibility compliance.

— The caveman skill reduces Claude token consumption by up to 65% by forcing terse, caveman-style English output.

— codebase-memory-mcp builds a persistent AST-based knowledge graph in pure C, claiming a 99% reduction in token usage for codebase understanding.

— Agent-Reach gives any AI agent zero-cost access to 15+ web platforms including Twitter, Reddit, GitHub, and YouTube through auto-routing Python modules.

— OpenMontage is the first open-source agent video production system, with 12 pipelines, 100+ tools, and 700+ composable skills.

— Vibe-Trading lets users describe trading strategies in natural language and executes the full research pipeline — data pulling, backtesting, and report generation.

— ai-hedge-fund simulates a 13-agent investment team with personas like Buffett, Munger, and Soros, each producing independent signals reviewed by a risk manager.

— destructive_command_guard is a Rust tool that intercepts dangerous agent commands like rm -rf, git push --force to main, and fork bombs.

— OfficeCLI wraps Word, Excel, and PowerPoint operations as CLI commands so agents can manipulate Office files without installing Microsoft Office.

— design.md from Google Labs merges YAML design tokens with Markdown specs into a single file that agents can read to generate visually consistent UI.

Conclusions

Agent Skills are undergoing the same standardization arc that containers experienced with Dockerfiles — moving from ad-hoc scripts to shareable, composable configuration files that define a new layer of the stack.

The simultaneous explosion of three competing Skills repositories with different philosophies (composable vs. methodological vs. quality-gated) suggests the ecosystem is still in a land-grab phase where no single approach has won.

Skills are decoupling from individual platforms; mattpocock/skills already works across Claude Code, Codex, Cursor, and Gemini CLI, hinting at an emerging cross-agent standard.

The caveman skill's 65% token reduction is a reminder that prompt engineering for cost efficiency is becoming as important as prompt engineering for accuracy.

codebase-memory-mcp's choice of pure C for a zero-dependency static binary signals that agent infrastructure is maturing to the point where performance and portability outweigh rapid prototyping convenience.

Agent-Reach's single-maintainer risk and lack of tests make it a textbook example of high-utility, high-risk infrastructure that teams should fork rather than depend on upstream.

OpenMontage's 700 composable skills mirror the Agent Skills pattern from coding tools, suggesting that skill modularity is becoming a cross-domain architectural pattern for AI agents.

Vibe-Trading and ai-hedge-fund represent two ends of the AI-finance spectrum: practical full-pipeline automation versus educational multi-agent simulation, and neither is yet a production trading system.

The appearance of destructive_command_guard as a dedicated safety tool indicates that agent terminal access is now common enough that guardrails are a distinct product category.

system_prompts_leaks accumulating 60,000 stars shows that developers treat reverse-engineered system prompts as valuable intelligence for building better application-layer prompts.

Concepts & terms

Agent Skills

Structured, reusable configurations that combine prompt templates, tool-calling definitions, and output format constraints to standardize how an AI coding agent performs a specific task. Analogous to a Dockerfile for agent behavior.

MCP (Model Context Protocol)

An open protocol that standardizes how AI agents connect to external tools and data sources, similar to how LSP (Language Server Protocol) standardized editor-language interactions. It replaces bespoke integrations with a universal interface for file systems, terminals, databases, and APIs.

AST (Abstract Syntax Tree)

A tree representation of the syntactic structure of source code. Tools like codebase-memory-mcp parse code into ASTs to build queryable knowledge graphs of function calls, type dependencies, and cross-file references.

Design Tokens

Platform-agnostic key-value pairs that represent design decisions like colors, spacing, typography, and border-radius. Stored in formats like YAML, they allow AI agents to generate visually consistent UI by referencing a single source of truth.

Worktree (Git)

A Git feature that allows multiple working trees attached to the same repository to be checked out simultaneously. Used by tools like orca and no-mistakes to isolate agent work and run verification in sandboxed directories before committing.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗