Someone just built a Rosetta Stone for AI coding assistants, and it changes the economics of who wins the IDE wars.
The Summary
- A new GitHub repo called wshobson/agents ships 84 plugins, 192 specialized agents, and 156 skill packages that work natively across Claude Code, Codex CLI, Cursor, OpenCode, GitHub Copilot, and Gemini CLI — from one markdown source
- Each IDE gets native-format artifacts, not watered-down translations, solving the "write once, runs everywhere badly" problem that's plagued developer tools since Java
- The repo demonstrates what happens when someone treats AI coding tools like package managers instead of monolithic black boxes
The Signal
The interesting part isn't the plugin count. It's the harness architecture. Most teams building for AI coding assistants pick one IDE and ship. Maybe they grudgingly support a second. This repo compiles from a single source tree into five different native formats, each optimized for how that specific IDE actually works.
That's not a technical flex. That's a business model unlock. If you're building AI-powered development tools, you've been facing a brutal choice: build deep for one IDE and hope it wins, or build shallow for many and watch developers ignore you because your tool feels like a web app pretending to be native. This approach suggests a third path: build once in a harness-agnostic format, compile to native idioms.
"One source-of-truth, five harnesses. Each harness gets idiomatic, harness-native artifacts — not lowest-common-denominator translations."
The plugin categories tell you what developers actually want from AI assistants:
- Domain experts for architecture, security, infrastructure, ML
- Scaffolding commands that generate boilerplate
- Security scanners that run in-editor
- Test generators that understand your stack
These aren't chatbot features. They're workflow automation that happens to use LLMs under the hood. The shift from "ask the AI a question" to "install the security plugin and run /scan" is the shift from Web2 copilot to Web4 agent.
The 192 agents break down into specializations that mirror how actual engineering teams structure: architecture agents, language-specific agents, infrastructure agents, security agents, data agents, ML agents, docs agents, business logic agents, SEO agents. That's not a random list. That's an org chart for an AI-native development team where each role is an installable module.
The Implication
Watch for two things. First, whether this multi-harness approach becomes table stakes for developer tooling. If you're building agents for programmers and you only support one IDE, you're betting that IDE wins and everyone else loses. That's a riskier bet than it was six months ago.
Second, watch how fast the "plugins as products" model spreads beyond coding. Plugins are just packaged agent capabilities with clear interfaces. Once developers normalize installing agent plugins the way they install npm packages, every other knowledge worker category will want the same thing. Figma plugins, but the plugins are AI agents. Notion blocks, but the blocks are specialized intelligences you can snap together.
The marketplace isn't the product. The compilation layer is.