Omniplug — AI skill for Claude Code

omniplug

Author an AI agent plugin **once** in a tool-neutral canonical format, then compile or install it into target-specific layouts. Claude Code, Cursor and Codex are supported today; future tools (Grok, Gemini CLI, …) slot in by implementing one adapter — no changes to the core.

Docs: **[asingamaneni.github.io/omniplug](https://asingamaneni.github.io/omniplug/)** — [Installation](https://asingamaneni.github.io/omniplug/docs/installation/) · [Usage](https://asingamaneni.github.io/omniplug/docs/usage/) · [Architecture](https://asingamaneni.github.io/omniplug/docs/architecture/)

Status

End-to-end pipeline (parse → IR → validate → compile → install) with three target adapters:

Target	Skills	MCP	Commands	Agents	Hooks	Guidance
claude	yes	yes	native	yes	yes	yes
cursor	yes	yes	rules	yes	yes	yes
codex	yes	yes	none (as skills)	no	yes	no

Both targets support every component natively. Where a canonical field has no native home, the adapter degrades it with a diagnostic instead of producing incorrect output — e.g. hook matchers are translated from Claude tool names to Cursor tool types (`Bash`→`Shell`, `Edit`→`Write`), a write-denying agent tool config becomes Cursor's `readonly: true`, and untranslatable matchers ship unfiltered with a warning rather than silently never firing. **Codex** compiles to a Codex plugin (`.codex-plugin/plugin.json`, `skills/`, `hooks/hooks.json`, `.mcp.json`), the unit `codex plugin add` installs from a marketplace. A Codex plugin has no slash commands, subagents or guidance file, so commands become skills and agents and guidance are dropped, each with a warning; hook scripts are rewritten to `${PLUGIN_ROOT}`, a stdio MCP server that runs a bundled file keeps its relative path and gets `cwd: "."` (Codex resolves it against the plugin directory and does not expand `${PLUGIN_ROOT}` in `.mcp.json`),