Docs

Skills

Install Convoy skills so AI clients work with your tasks in a consistent way.

Skills teach supported AI clients (Claude Code, Codex) how to work with Convoy: standard task workflows, artifact conventions, and less repeated setup across teammates. They complement MCP — MCP exposes the tools, skills teach the client how to use them well.

Install with the CLI

convoy init skill
convoy init skill --cli claude,codex --scope project
convoy init skill --cli claude,codex --scope project --force

Project-scoped installs keep one copy of the skill in .agents/skills/convoy-tasks. Claude Code raw skill installs create .claude/skills/convoy-tasks as a symlink to that shared copy. If an existing skill directory or link is found, Convoy prompts before replacing it; use --force to replace without a prompt. The --claude-method flag selects the Claude Code install method when more than one is available.

Scope choices

  • use project scope when the skill should live with the repo — preferred for shared team repos
  • use user scope when the setup should stay local to one machine

skills.sh alternative

If your team already manages agent skills with skills.sh, install the same task skill with:

npx skills add https://github.com/packtrain/convoy/tree/main/plugins/convoy/skills --skill convoy-tasks

The /tree/main/plugins/convoy/skills form is the documented GitHub subdirectory URL shape for skills.sh — a repository URL without /tree/main/ points at the repository root, not this skill directory.

Useful options:

  • --agent claude-code or --agent codex targets a specific client
  • --global installs to the user-level skill directory
  • --copy copies files instead of using skills.sh's default symlink-based install
  • --yes skips confirmation prompts

See the skills CLI reference and vercel-labs/skills for the full command surface.

  1. Save the Convoy profile (convoy setup).
  2. Generate MCP config (Set up MCP).
  3. Install the project-scoped skill (convoy init skill --scope project).
  4. Validate the client inside that repo: ask it to list tasks, then create a test task.

On this page