skill-cli

One skill store. Clean agent folders. Activate from anywhere.

skill-cli is a cross-agent skill manager for AI coding agents (Claude Code, Codex, Cursor, Gemini, …). Skills live in a single global store — your agent directories stay clean — and you control which skill is active where with one config file.

• 🗂️ One canonical store — ~/.skill-cli/store. No more copies scattered across ~/.claude, ~/.codex, ~/.cursor, …
• 🧹 Agent folders untouched — nothing is written into agent directories.
• 🎛️ Three-layer activation — installed → enabled globally → enabled per project. allow always wins over deny.
• 🔌 Universal sources — owner/repo, GitHub/GitLab URL, git URL, local path, npm package (via npx skills).
• ⚡ Pull-based — agents pull skill content into context on demand via skill trigger /X. No hooks required.
• 🔍 Interactive TUIs — skill search to discover & multi-install from the skills registry; skill (no args) to manage every installed skill with the keyboard. TTY-only; agents & CI stay non-interactive.
• ⭐ Default skills — one global list: active by default in every project AND auto-loaded on session start.
• 🖥️ Cross-platform — Windows, macOS, Linux (handles the Windows npx spawn quirk for you).
• 🧪 Well tested — 225 unit + CLI tests, network-free by default.

⭐ Found this useful?

If skill-cli saves you time, please give it a star ⭐ on GitHub — it helps others discover the project and keeps development going. Thank you! 💜

Installation

npm install -g @victortomaili/skill-cli

Requires Node.js 22+.

Quick start

# 1) one-time global setup (creates the store + bootstraps detected agents)
skill init -g

# 2) install a skill from any source (or `skill search` to browse interactively)
skill install owner/repo

# 3) manage everything with the keyboard (terminal only)
skill              # ↑↓ move · space toggle · a default · d delete · enter view
skill list
skill cat <name>

That's it. Your agent now knows: when the user types /X, run skill trigger X.

How it works

~/.skill-cli/
  store/<skill>/SKILL.md     ← one canonical store (skills live here)
  config.yaml                ← global defaults (active by default + auto-load)

<project>/skill.config       ← per-project overrides (inherit / deny / allow)
~/.claude/CLAUDE.md          ← bootstrap block injected by `skill init -g`
~/.codex/AGENTS.md             (idempotent — preserves your existing content)
~/.gemini/GEMINI.md
~/.pi/agent/AGENTS.md

Agents don't read skill.config directly — the CLI does. The agent only needs the short bootstrap block telling it to run skill on /X.

Commands

Command	Description
skill init -g	Global setup: create store + inject bootstrap into agent files
skill init	Create a skill.config for the current project
skill install <source>	Fetch skill(s) to the store (agent dirs untouched)
skill search	Interactive search & multi-install from the skills registry (TTY)
skill / skill manager	Interactive manager: toggle, default, delete, view (TTY)
skill enable <name> [-g]	Allow in project, or global default with -g
skill disable <name> [-g]	Deny in project, or remove global default with -g
skill default <name>	Mark a default skill (global: active + auto-load)
skill undefault <name>	Remove the default flag
skill active	List ACTIVE skills + descriptions (the command your agent runs on start; alias: status)
skill list	Show installed + active skills (cwd-aware, ★ = default)
skill show <name>	Skill metadata (path, triggers, version)
skill cat <name>	Dump skill content into context
skill trigger <keyword|name>	/X trigger or skill name → content (single), candidates (multi)
skill update [name…|--all]	Re-fetch from source, update changed skills
skill remove <name> [-y]	Remove from store (prompts on TTY; agents / CI / -y skip)

Aliases: ls (list), add (install), info (show), rm/uninstall (remove), browse (search), ui (manager), def (default), undef (undefault).

Install sources

install delegates resolution to npx skills, so every format works:

owner/repo                                        GitHub shorthand
https://github.com/owner/repo                     full GitHub URL
https://github.com/owner/repo/tree/main/skills/x  path inside a repo
https://gitlab.com/org/repo                       GitLab URL
git@github.com:owner/repo.git                     any git URL
./my-local-skills                                 local path
<npm-package>                                     npm package shipping skills

Install runs npx skills add in a temp cwd so files land in <tmp>/.claude/skills/ and are then moved into ~/.skill-cli/store/. Your real agent directories are never written to.

Skill format

Faithful to the SKILL.md standard (the author decides the structure):

---
name: deep-research
description: Deep source-research workflow
version: 1.2.0
triggers: [research, deep-search]   # skill-cli-specific; powers /<trigger>
---
# Your instructions here

triggers is optional and skill-cli-specific. See examples/hello-world/SKILL.md for a minimal, copy-paste template.

Activation

One global concept, refined per project:

1. installed — present in the store (passive)
2. default (global) — listed in config.yaml defaults:; active by default in every project and auto-loaded on agent session start. Mark one with skill default <name> (or skill enable -g).
3. per-project override — a project's skill.config refines the set:
   • allow — activate an otherwise-passive skill in this project only
   • deny — turn off a default skill in this project only
   • inherit: false — ignore global defaults entirely (only allow applies)

A project with no skill.config inherits the global defaults (all active). Pure allowlist mode in a single project:

# <project>/skill.config
inherit: false                 # ignore global defaults in this project
allow: [react-best-practices]  # only this skill is active here

allow always wins over deny, so with inherit: true you can also write deny: ["*"] + allow: [X] = "only X here". Matching is case-insensitive throughout.

Interactive mode (terminal only)

When run in a real terminal, skill-cli offers two keyboard UIs. Agents and CI never enter them (a non-TTY stdin is detected) — they use the plain commands.

skill — manage installed skills

Opens a full-screen manager over your store:

Key	Action
↑ / ↓	move
space	toggle active in the current project
a	toggle the global default (active + auto-load)
d	delete (asks y/N)
enter	view the SKILL.md
q / esc	quit

skill search — discover & install

Type a query → browse live results from the skills registry (npx skills find) → mark one or more with space → enter installs all marked → loops back so you can search again. esc or an empty query quits. skill install with no source also opens it.

Default skills

A default is a skill that is active by default in every project AND auto-loaded on agent session start — one unified global list in config.yaml (defaults:). Mark one with skill default <name>, remove with skill undefault <name>, list them with skill active. The AGENTS.md bootstrap block tells your agent to run skill active on start, then skill cat <name> for each.

Defaults are a global concept (never per-project). A project can still turn a default off locally with skill disable <name> (a project deny), or activate a non-default locally with skill enable <name> (a project allow) — see Activation. In the interactive manager, a toggles the global default; space toggles the per-project override.

Bootstrap

skill init -g injects a short, idempotent block into each detected agent's global instruction file (CLAUDE.md, AGENTS.md, GEMINI.md):

START GATE: on the first user message, your VERY FIRST action — BEFORE ANYTHING ELSE (before thinking, before any tool call) — is skill active, the skill catalog (every active skill's name + full description; never the body). Read it and decide per skill from its description: AUTO-LOAD (silently) anything needed to do the task at all, or that makes the work higher-quality / more correct (domain knowledge, best-practice, standards). PROPOSE (apply only after the user confirms; /X = confirm) anything that's an optional change to the experience — response style/format/mode, context-cost optimization, or an optional shortcut (easier/faster, not about correctness). When unsure, propose. LOADED ≠ LISTED: a skill is loaded only if you cat-ed it this session. PRIORITIES: correctness > quality > cost (cheap) > speed — never sacrifice correctness/quality for speed or cost. /X → skill trigger X.

It's wrapped in <!-- BEGIN skill-cli --> … <!-- END skill-cli --> markers, never duplicates, and preserves your existing file content. Re-run init -g any time — it reports already set up and rewrites nothing.

Supported agents: Claude Code, Codex, Gemini, and pi. (Cursor uses .cursor/rules with a different format — adapter planned. See issue tracker.)

Cross-platform

Works on Windows, macOS, and Linux. On Windows it spawns npx via cmd.exe /c (the Node 20+ CVE-2024-27980 workaround) with shell:false and validates the source against shell metacharacters, so it's safe by default.

Development

git clone https://github.com/victortomaili/skill-cli
cd skill-cli
npm install
npm link            # makes `skill` point at your checkout
npm test            # 225 tests, network-free (~3s)
npm run test:e2e    # opt-in: real npx fetch over the network

Tests never touch your real ~ — they use an isolated SKILL_CLI_HOME. See CONTRIBUTING.md for the project structure, how to add a command, and the PR checklist.

Contributing

Contributions are welcome! 💜 Please read CONTRIBUTING.md and follow our Code of Conduct.

• 🐛 Report a bug
• ✨ Request a feature
• 🔒 Found a security issue? See SECURITY.md — report it privately, not as a public issue.

Changelog

See CHANGELOG.md.

License

MIT © skill-cli contributors