Skills is a command-line tool from Vercel Labs for the open agent skills ecosystem. An agent skill is a directory holding a SKILL.md file with YAML frontmatter (a name and a description) followed by the instructions the agent follows when the skill is activated, optionally with bundled scripts and reference files. The CLI installs, lists, searches, updates and removes such skills across 79 coding agents, including OpenCode, Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Cline, Roo Code, Goose, Hermes Agent, OpenClaw, Kilo Code, Qwen Code, Windsurf, Droid and Antigravity.
A skill source can be a GitHub owner/repo shorthand, a full GitHub, GitLab or Azure Repos URL, a direct path to one skill inside a repository, any git URL, a local path, or a direct download URL pointing at a SKILL.md file or a zip, tar, tar.gz or tgz archive. Private repositories reuse the Git credentials, GitHub CLI login or SSH key already configured on the machine, and the CLI never copies a stored GitHub CLI credential into its own process. Skills are installed at project scope (committed with the repository) or global scope, either as symlinks to one canonical copy or as independent copies per agent. The CLI detects which agents are installed and otherwise prompts for targets.
Inside a repository the CLI discovers skills at the root, under a skills directory and its curated, experimental and system subfolders, and in every supported agent's own skills folder, walking up to three levels deep so catalog layouts with one or two category levels are found. Skills declared in a Claude Code plugin marketplace or plugin manifest are discovered as well. A use command generates a prompt for one skill without installing it, or starts a supported agent interactively with that prompt. Skills follow the shared Agent Skills specification, with some agent-specific features: allowed-tools is supported by most agents, hooks by Claude Code, Cline and Kiro CLI, and context: fork by Claude Code only. Anonymous usage telemetry can be disabled.
Features
- Install from many sources: GitHub shorthand, full GitHub, GitLab or Azure Repos URLs, a path to one skill in a repo, any git URL, a local directory, or a direct SKILL.md or archive download URL
- 79 supported agents: including Claude Code, Codex, OpenCode, Cursor, Gemini CLI, GitHub Copilot, Cline, Roo Code, Goose, Hermes Agent, OpenClaw, Kilo Code, Qwen Code, Windsurf, Droid and Antigravity, with per-agent project and global paths
- Private repositories: reuses configured Git credentials, GitHub CLI login or SSH, with optional explicit GitHub tokens for API access
- Project or global scope: skills live with the project for the team or in the user directory for all projects
- Symlink or copy: one canonical copy symlinked into each agent for easy updates, or independent copies where symlinks are unsupported
- Use without installing: generate a prompt for a single skill, or start a supported agent interactively with it
- Lifecycle commands: list installed skills, search interactively or by keyword (optionally across one owner's repositories), update by name or scope, remove by name, source or agent, and scaffold a new SKILL.md
- Skill discovery: root, skills folders and every agent's skills directory, walked three levels deep, plus skills declared in Claude Code plugin manifests, with a full-depth option
- Non-interactive installs: flags to pick agents and skills, skip prompts and install everything, for CI pipelines
- Internal skills: a frontmatter flag hides work-in-progress skills from discovery unless explicitly enabled
- Download limits: archives are capped at 10 MiB downloaded, 25 MiB extracted and 1000 files by default, with overrides for trusted sources
