# Skills, Krew tools, and MCP

Three layers, deliberately separate: a SKILL changes what the model knows for a turn; a TOOL changes what it can do; an MCP SERVER changes what tools exist.

## Skills teach
Markdown documents (checklist, runbook, house style) loaded on demand — models see names + one-liners, load content only when relevant. Locations (earlier wins on clash): project .kaptain/skills > machine-wide library > kapsule > imported packs. https://chynj.ca/kaptain/docs/how-to/skills/

## Krew tools act
Built-in belt: KodeGraph code intelligence, files, shell/tests/git, web research, delegation to local models, memory, agent-to-agent. Per-tool switches in the Krew view. The same belt serves outward to Claude/Codex/Gemini CLIs as the `kaptain-engine` MCP server.

## MCP extends
Your tool = a program speaking MCP: connect as a server (UI or `kaptain mcp add`), ship inside a kapsule, or drop an executable into plugins/. Same catalog, same switches. Deliberately no paste-a-shell-script path — tools are real programs behind a real protocol.
https://chynj.ca/kaptain/docs/how-to/mcp/ · https://chynj.ca/kaptain/docs/how-to/krew-tools/

## One gate
Every risky call passes the same four-mode approval (Ask / Read-only / Auto in workspace / Full), per model or per kapsule; trusted scopes pre-approve targets, never whole tools.