Agent skill

mxkey ships with a Claude Code / Cursor / Codex skill (SKILL.md plus the references/ directory) that loads when present in your editor's skills directory. With the skill active, the agent recognises secret-related situations and runs mxkey for you — without ever seeing the value.

Install

Drop the cloned repo into your editor's skills folder:

# Claude Code
ln -sfn "$(pwd)" ~/.claude/skills/mxkey

# Cursor (path may differ)
ln -sfn "$(pwd)" ~/.cursor/skills/mxkey

# Codex (check your editor's docs)

A symlink so git pull updates the skill in place.

The agent also expects the mxkey CLI on PATH, so make sure bash install.sh has run.

What it does

How the skill is structured

The repo contains:

mxkey/
├── SKILL.md            ← the main skill instructions (160 lines)
├── references/
│   ├── setup.md
│   ├── migrate-from-env.md
│   ├── troubleshooting.md
│   └── keychain-deep-dive.md
└── mxkey               ← the CLI binary the skill calls

SKILL.md is YAML-frontmattered and Markdown-bodied. The agent reads it on session start; the references/ files are loaded on demand when the relevant playbook fires.

Why a skill, not a server

The skill is just text — it doesn't run code. The agent uses it as context-on-demand: when you mention an API key, Claude Code reads SKILL.md, sees the playbook for "user mentions an API key", and runs the corresponding mxkey set/mxkey run commands itself.

There's nothing to authenticate, no daemon, no MCP server. The mxkey CLI is the only moving part — the skill is documentation that the agent reads.

Versioning

The skill version (in SKILL.md frontmatter) is bumped alongside the CLI. Agents keep the symlinked version automatically when you git pull.

See also

• Install — CLI install
• Hard rules — what the agent will refuse to do
• GitHub: SKILL.md source