A package manager for AI coding agent config sprawl

What it does

CAPA treats AI coding agent configuration as code. A single capabilities.yaml declares skills, tools, rules, sub-agents, MCP servers, and plugins; the tool transpiles those declarations into each agent’s native format—whether that’s .cursor/rules/ for Cursor, .claude/agents/ for Claude Code, or AGENTS.md for Codex. A capabilities.lock file pins resolved SHAs so teammates get identical bytes on the next clone.

The interesting bit

Instead of front-loading every tool into every agent, CAPA registers one MCP endpoint per agent and lazy-loads tools on demand. The README claims this cut inference costs by 19–40% across 150 trials on claude-opus-4-8, while filtered endpoints ensure a research sub-agent never accidentally sees a git push tool it shouldn’t touch.

Key highlights

• One capabilities.yaml drives 35+ agents, including Cursor, Claude Code, Codex, Windsurf, and GitHub Copilot.
• Hand-written content in target files is preserved via capa-managed marker blocks.
• Sub-agents get their own filtered MCP endpoints, so they only see relevant tools.
• Every defined tool is exposed as a native CLI command under capa sh.
• A bootstrap skill can scan an existing project and pull scattered rules into CAPA management.

Verdict

Worth evaluating if your team is juggling multiple AI agents and config drift between editors. If you live inside a single editor with one rules file, it’s likely overkill.