Agent configuration is scattered across CLAUDE.md, .cursor/rules, AGENTS.md, MCP JSON, hooks and skill folders. No two teammates match, nothing is pinned, and cloning the repo does not clone the setup. CAPA collapses all of it into one file next to your code.
- One file, 35+ agents
- SHA-pinned in capabilities.lock
- MCP gateway built in
- MIT licensed
One file, every agent
Write a capability once and CAPA fans it out to each provider's native layout. Marker blocks mean it writes surgically, leaving anything you hand-edited alone.
- Skills, tools, rules, sub-agents, MCP servers, hooks and plugins in one manifest
- Native output for .cursor/rules, .claude/agents, AGENTS.md and the rest
- capabilities.lock pins SHAs, so tomorrow's clone gets today's bytes
An MCP gateway on your laptop
At runtime CAPA is one local MCP endpoint per project. It proxies stdio, HTTP and SSE servers, loads tools on demand instead of front-loading the catalogue, and filters what each sub-agent can call.
- On-demand tool loading cut token use 19 to 40 percent across 150 trials
- Each sub-agent gets a filtered endpoint, so a research agent never inherits git push
- A live activity trace of every MCP call, shell tool and agent span
Run agents without letting them touch your repo
capa wrap runs Cursor, Claude or Codex from a shadow workspace, so provider directories never land in your real working tree. A local web UI edits the manifest with live YAML sync, registry browsing and OAuth setup.
Pricing
Free and open source under the MIT licence. macOS, Linux and Windows.
