Skip to main content

Plugins & Skills

ak-coder can be extended in two ways:

ExtensionWhat it addsWhere
PluginNew tools the LLM can call.ak-coder/plugins/<name>/
SkillNew slash commands (/mycommand)Any SKILL.md in the workspace

Plugins (MCP tools)​

Plugins run as local MCP (Model Context Protocol) servers — spawned as child processes, communicating over stdio JSON-RPC. At startup, ak-coder discovers every plugin.json under .ak-coder/plugins/ and registers the plugin's tools as serverName__toolName.

Getting started: Building a Plugin

Key points:

  • Use @ak-coder/sdk and registerTool() — install SDK via file:…/packages/sdk from a local checkout; see Building a Plugin
  • Never write to stdout in plugin code — it's the JSON-RPC transport
  • Optional outputSchema validates handler output (warns on mismatch, does not abort)
  • Plugin tools do not yet support tool annotations — they always run sequentially
  • Lifecycle hooks in .ak-coder/hooks/ can intercept writes and commands from core tools (not plugin internals)

See ADR 03: Plugin System & Hooks and ADR 07: Plugin Output Schema.

Skills (slash commands)​

Skills are markdown instruction files loaded at startup and on reload. Typing /skills:<name> injects the skill body (plus any arguments) into the agent context.

Getting started: Skills

Key points:

  • Any file named SKILL.md anywhere in the workspace is discovered automatically
  • Front-matter name drives the slash command; description appears in /help and tab completion
  • /skills reload rescans; editing a SKILL.md via write tools also triggers reload
  • Skills are instructions for the LLM — they do not add new executable tools
  • Tab completion: type /skills: + Tab to pick a loaded skill

See ADR 04: Skills System.