Concept · Capability · 4 min read

Skills, Krew tools, and MCP

Everything a Kaptain model can do beyond talking comes from three mechanisms with deliberately different shapes: skills (knowledge it can load), Krew tools (actions it can take), and MCP (the door for tools you bring).

Three layers, three jobs

A skill changes what the model knows for a turn. A tool changes what it can do. An MCP server changes what tools exist at all. Collapsing these into one mechanism is how systems end up with prompt files that execute and plugins that secretly prompt — Kaptain keeps them separate so each can be reviewed on its own terms.

Skills teach

A skill is a Markdown document — a review checklist, a deploy runbook, your house style — that models pull into context exactly when relevant. They see the list (names + one-liners) and load content on demand, so a big library costs nothing until used. Skills live in the project (.kaptain/skills), your machine-wide library, a kapsule, or imported packs — earlier sources win on a name clash. The pool also lists the skills your Claude and Codex CLIs already have (their project and machine-wide folders), each shown with its source. How-To ▸ Skills.

Krew tools act

The built-in belt covers code intelligence (KodeGraph), files, shell/tests/git, web research, delegation to local models, memory, and agent-to-agent calls — every tool with its own on/off switch in the Krew view. The same belt is served outward to Claude/Codex/Gemini CLIs as the kaptain-engine MCP server, so your terminal assistants get Kaptain's hands too.

MCP extends

Your own tool is a program that speaks MCP — connected as a server (UI or kaptain mcp add), shipped inside a kapsule, or dropped as an executable into plugins/. However it arrives, it lands in the same catalog with the same switches. There is deliberately no paste-a-shell-script tool path: a tool is a real program behind a real protocol, so it can be versioned and reviewed. How-To ▸ MCP · How-To ▸ Krew tools.

One gate over all of it

Built-in or brought-in, every risky call passes the same four-mode approval — Ask, Read-only, Auto-approve in workspace, Full access — set per model or per kapsule. A kapsule can pre-trust specific targets (one path, one server, one host) without widening anything else.

Agent-readable version: index.md