Things I've built, mostly so I'd want to use them.

Define an agent once in a shared source tree and generate consumer-native files for each harness: agent definitions for Claude Code, GitHub Copilot, Codex, Cursor, and Gemini CLI, plus skill bundles for Claude Skills, neutral Agent Skills, and Hermes. custom_agents absorbs each runtime’s file layout, prompt format, and model conventions so one canonical agent.yaml ships everywhere.

Each definition picks an export mode (agent, skill, or none) and can declare per-agent harness include/exclude preferences. The shared-agents CLI handles sync, list, validate, and clean with comma-separated harness/agent filters, an interactive TUI for picking targets with a dry-run preview, and manifest-based cleanup so it only ever removes files it generated.