agents-skills

English | 简体中文

A minimal installer and manager for AI agent skills: all skills live in one canonical directory, and a single agent --link makes them visible and usable to Claude Code, Codex, Cursor, and 70+ other coding agents — install once, works everywhere.

cargo install agents-skills

Also ships as an embeddable Rust library — see docs/LIBRARY.md.

Quick start

agents-skills agent --link            # link all installed agents
agents-skills add anthropics/skills   # install a skill pack (visible to every agent immediately once it lands in the canonical directory)
agents-skills list                    # list installed skills
agents-skills agent --status          # show the link status of each agent

The core idea is linking: skills are stored exactly once in the canonical directory (project-level .agents/skills/ or global ~/.agents/skills/), and agent --link creates, for every installed agent, a symlink in its skills directory pointing at the canonical directory; skills installed afterwards via add become visible to all agents immediately — no syncing needed.

How it works

add/remove/update/disable/enable only operate on the canonical directory; agents share it automatically through the symlinks. Common commands:

agents-skills agent --link claude-code            # link (pre-existing content is backed up automatically)
agents-skills agent --link claude-code --migrate  # link and migrate pre-existing skills into the canonical directory
agents-skills agent --unlink claude-code          # unlink (and restore the backed-up content)
agents-skills add anthropics/skills@pdf            # install only the specified skill
agents-skills list --json                          # machine-readable output
agents-skills remove pdf                           # remove a skill
agents-skills update                               # update to the latest versions per the lockfile
agents-skills disable pdf                          # disable (moved out of the canonical directory, files kept)
agents-skills enable pdf                           # re-enable (inverse of disable)

• Commands operate on the global scope ~/.agents/skills by default; --project <dir> switches to the project scope (.agents/skills under the given directory; use --project . for the current directory).
• When an agent's skills directory already has content, agent --link does not refuse: the whole directory is moved as-is into the backup slot .agents/backup-skills/<agent>/skills/, and agent --unlink restores it in full; with --migrate, the skills inside are moved into the canonical directory (on name conflicts the canonical copy wins, and the agent-side copy stays in the backup).
• For unlinked agents, agent --status categorizes the contents of their own skills directory (private skills / other files) plus any backup waiting to be restored (backup parked at); content of linked/canonical agents is shown by list.
• update skips disabled skills; list always shows every skill, tagged enabled/disabled.

Source formats

The <source> argument of add accepts:

Within a repository, skills are discovered in priority-ordered container directories (skills/, .curated/, .experimental/, .system/); shallower directories shadow deeper ones.

Install locations

• Canonical directory (the single real copy) — project-level ./.agents/skills/<name>, global ~/.agents/skills/<name>.
• Agent integration — agents that do not natively read the canonical directory get a directory-level symlink: .claude/skills → ../.agents/skills (project) or ~/.claude/skills → ~/.agents/skills (global).

Command cheat sheet

Commands have no aliases (a minimal interface — full names only).

For the full command-line reference see docs/CLI.md; library users see docs/LIBRARY.md; project developers see docs/DEVELOPER.md.

License

Licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.