glab-skills

Claude Code skills and hooks for working with GitLab via the glab CLI.

What's in here

This repo has four skills covering common GitLab operations: auth, merge requests, issues, and CI pipelines. It also includes glab-auth-guard.sh, a PostToolUse hook that detects authentication errors in glab output and prompts you to log in, and a settings.local.json with pre-approved glab command permissions so Claude does not ask for confirmation on every call. 224 commands are documented across docs/commands.md and docs/reference.md.

Requirements

• glab CLI installed
• Claude Code installed

---

Install glab

macOS:

brew install glab

Windows (winget):

winget install GitLab.GLab

Windows (scoop):

scoop install glab

Linux (apt):

sudo apt install glab

Linux (snap):

sudo snap install glab

Verify the install:

glab --version

Authenticate

You must authenticate before any glab command will work.

glab auth login

For a self-managed GitLab instance:

glab auth login --hostname gitlab.example.com

Check your current auth status:

glab auth status

The glab-auth-guard.sh hook runs after every Bash tool call. When it detects an authentication error in glab output, it tells Claude to prompt you to run glab auth login.

---

Plugin install (recommended)

Run these three commands inside any Claude Code session:

/plugin marketplace add moumine9/glab-skills
/plugin install glab-skills
/reload-plugins

Pick a scope when prompted. User is the one you want most of the time — it's available in every project. Project drops the plugin into .claude/plugins/ so the whole team gets it. Local keeps it project-scoped and just for you.

Skills are available as:

• /glab-skills:auth
• /glab-skills:mr
• /glab-skills:issue
• /glab-skills:ci

The auth guard hook activates automatically once the plugin is installed.

Note: The official Anthropic marketplace includes a gitlab plugin that connects via MCP. This plugin is different: it wraps the glab CLI using skills, which load progressively and use far less context than MCP tool definitions loaded at session start.

Test locally before installing

claude --plugin-dir ./

This loads the plugin from the current directory without installing it. Use /reload-plugins to pick up changes without restarting.

Manual install

Skills

Copy each skill directory into ~/.claude/skills/:

cp -r skills/auth ~/.claude/skills/
cp -r skills/mr ~/.claude/skills/
cp -r skills/issue ~/.claude/skills/
cp -r skills/ci ~/.claude/skills/

Standalone triggers (without the plugin namespace):

• /auth
• /mr
• /issue
• /ci

Auth guard hook

Copy the hook and make it executable:

cp hooks/glab-auth-guard.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/glab-auth-guard.sh

Add the following entry to ~/.claude/settings.json under hooks.PostToolUse:

{
  "matcher": "Bash",
  "hooks": [
    {
      "type": "command",
      "command": "bash \"${CLAUDE_CONFIG_DIR:-$HOME/.claude}/hooks/glab-auth-guard.sh\""
    }
  ]
}

Project permissions

Copy .claude/settings.local.json into your project's .claude/ folder:

cp .claude/settings.local.json <your-project>/.claude/settings.local.json

---

Reference

• docs/commands.md: every command at a glance
• docs/reference.md: the full reference, flags and examples included
• CHANGELOG.md: plugin version history

Skills

Skill	Plugin trigger	Standalone trigger	Description	Auto-invoked
auth	/glab-skills:auth	/auth	Auth flows, Docker credentials, tokens, and key management	No
mr	/glab-skills:mr	/mr	Full MR lifecycle: create, review, merge, and stacked diffs	No
issue	/glab-skills:issue	/issue	Create, manage, board view, and track issues	No
ci	/glab-skills:ci	/ci	Pipelines, jobs, schedules, runners, and CI variables	No

Skills with disable-model-invocation: true require explicit invocation. Claude will not trigger them automatically.

Hooks

Hook	Trigger	What it does
glab-auth-guard.sh	PostToolUse + PostToolUseFailure (Bash)	Detects auth errors in glab output and tells Claude to prompt for authentication

The hook uses an if: "Bash(glab *)" condition so it only spawns when a glab command is involved, not on every Bash call.

Settings

.claude/settings.local.json contains pre-approved permissions for all glab subcommands (auth, mr, issue, ci, repo, config) and common git operations, so Claude can run them without asking for confirmation each time.