superskill

Multi-agent skill, command, subagent, magent, hook, and MCP config distribution and authoring CLI. Install a Claude Code plugin to any coding agent, and create / validate / evaluate / refine / evolve agent-facing content from the command line.

Install

# go with bun, suggested
bun add -g @gobing-ai/superskill

# or, run ad-hoc, no install:
bunx @gobing-ai/superskill --help

# or, go with npm
npm i -g @gobing-ai/superskill

No clone, no checkout, no extra flags. The published package bundles the cc plugin and the team-stark-children main-agent config, and the CLI locates them at runtime, so install works from any directory:

# go with owner/repo shorthand → GitHub
superskill install cc --magent team-stark-children --marketplace gobing-ai/superskill

# or, go with the full path of the github URL
superskill install cc --magent team-stark-children --marketplace https://github.com/gobing-ai/superskill

# or, go with pre-installed via bun
superskill install cc --magent team-stark-children --marketplace $(dirname $(bun pm bin -g))/install/global/node_modules/@gobing-ai/superskill

# or, go with pre-installed via npm
superskill install cc --magent team-stark-children --marketplace $(dirname $(npm root -g))/node_modules/@gobing-ai/superskill

# or, go with already downloaded source code
superskill install cc --magent team-stark-children --marketplace /Users/robin/xprojects/superskill

That distributes cc (skills, commands, subagents, hooks, rules) to every configured target and emits the team-stark-children magent. Want only the plugin, or only specific agents? superskill install cc (no --magent) and superskill install cc --targets codex,pi work too. See cmd_install.md for the full surface, including installing from a GitHub marketplace (--marketplace gobing-ai/superskill). For native Claude Code marketplace installation (claude plugin install cc@superskill), see the Native cc Marketplace Pilot.

From source (contributors):

proto use          # install pinned tool versions (Bun, etc.)
bun install
bun run build
cd apps/cli && bun link

Requires Bun ≥ 1.4.3. See Installation guide for details and troubleshooting.

Quick start

# Install the bundled cc plugin to every supported target
superskill install cc --targets all

# Install cc plus the team-stark-children main-agent config
superskill install cc --magent team-stark-children

# Author a skill: scaffold → validate → evaluate → refine
superskill skill scaffold my-skill --description "Deploy a Cloudflare Worker"
superskill skill validate my-skill
superskill skill evaluate my-skill --save
superskill skill refine my-skill --auto

# Package a skill for distribution, or merge skills
superskill skill package my-skill
superskill skill migrate ./source-skill-1 ./source-skill-2 ./dest-skill

# Emit a hook to a single target agent
superskill hook emit my-hook --target pi

Full walkthrough: Quick start guide.

Supported agents

| Agent | Skills | Commands | Subagents | Hooks | | ------------------- | :----: | :------: | :-------: | :---: | | Claude Code | ✓ | ✓ | ✓ | ✓ | | Grok | ✓ | ✓ | ✓ | ✓ | | Codex | ✓ | ✓ | ✓ | ✓ | | Pi | ✓ | ✓ | ✓ | ✓ | | omp | ✓ | ✓ | ✓ | ✓ | | OpenCode | ✓ | ✓ | ✓ | ✓ | | Antigravity IDE | ✓ | ✓ | — | ✓ | | Antigravity CLI | ✓ | ✓ | — | ✓ | | Hermes | ✓ | ✓ | — | ✓ | | Grok Bot (opt-in) ⁶ | ✓ | — | — | — |

Both Antigravity targets get skills, commands, and hooks from rulesync (codexcli/antigravity-cli/antigravity-ide adapters); subagents are not part of the rulesync→antigravity map, so they don't ship there. omp, Grok, and Claude Code install plugins natively (no rulesync pass); the remaining targets route through rulesync's per-target generators. Hermes copies skills/commands from OpenCode's rulesync output and adds a canonical hooks.json copy at ~/.hermes/hooks.json.

Pi loads extensions natively from a plugin's plugin.json extensions.pi array (v0.3.12+), installing them to .pi/agent/plugins/<plugin>/ with generated package.json and settings.json registration. Plugins without extensions.pi get their hooks.json hooks as generated native Pi extension modules (~/.pi/agent/extensions/superskill-<plugin>-<hook>.ts): deny blocks the tool call, ask opens Pi's confirm dialog, and a Stop block continues the run. No @vahor/pi-hooks dependency (ADR-039).

Grok loads the Claude-format plugin natively — slash form is /plugin:command (e.g. /cc:skill-add). Codex/Pi-installed skills under ~/.agents/ may also appear in Grok as /plugin-command (hyphen form) when Grok discovers the shared skills root; prefer the colon form for plugin commands.

⁶ Grok Bot (Grok chat on its own VPS/Sand host) is an opt-in, install-only target (--targets grok-bot, excluded from all; ADR-036). It publishes the flat skill catalog to the Sand workflows/ directory (bridge by default: thin pointers to a private canonical copy under <sandRoot>/.superskill/grok-bot/skills/; --materialize full for everything under workflows/). Slash form is /plugin-skill-name; commands/subagents install as skills/playbooks, hooks/MCP are not installed. Every install/update also prepares a deterministic registration handoff at <sandRoot>/.superskill/grok-bot/register/<plugin>.json (task 0130) — slash registration is NOT automatic: filesystem install does not fill chat /. Have the host agent consume the handoff with a verified host method; picker visibility and enablement stay unverified until observed. First use with an empty slash picker: ask a Bot to read and follow <sandRoot>/workflows/cc-grok-bot-register/SKILL.md (bundled cc recovery skill, task 0132) with --plugin <installed-plugin> — or, when cc is absent, paste the validated handoff recovery prompt printed by install; bridge's hidden .superskill recipes may need the Bot's authorized Shell tool if Read denies them (do not rewrite recipes onto ~/.agents/skills). Host registry upserts may rewrite bridge workflow pointers; that alone is not owned-drift when canonical hashes still match. Run superskill doctor --targets grok-bot for a filesystem health check (it reports slash-registry status as unknown with next steps).

See entity locations for the exact install directories per agent.

Agents that don't natively support some entity types still get them. superskill install adapts commands and subagents as Skills 2.0 skill directories for targets that lack them — so every agent receives the full plugin surface, regardless of native feature set.

Commands

| Command | What it does | Docs | | --------- | -------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------ | | install | Distribute a plugin's skills, commands, subagents, magents, hooks, and MCP config to target agents | cmd_install.md | | agent | Manage subagent definitions (scaffold / validate / evaluate / refine / evolve) | cmd_agent.md | | skill | Manage skill definitions (lifecycle + package, migrate) | cmd_skill.md | | command | Manage slash command definitions | cmd_command.md | | hook | Manage hook definitions (+ emit, run) | cmd_hook.md | | magent | Manage main-agent configurations | cmd_magent.md | | script | Run / resolve / build portable twins for plugin scripts (run, path, convert) | how to organize | | doctor | Read-only health check for the grok-bot Sand target (--targets grok-bot [--json]) | entity locations |

The five type commands share a common lifecycle: scaffold → validate → evaluate → refine → evolve, with type-specific quality dimensions and rubrics.

Further reading

| Topic | Document | | ---------------------------------------------- | -------------------------------------------------------------- | | Full help index | docs/help/index.md | | Quality system (rubrics, scoring, evolve gate) | docs/help/quality_system.md | | Entity locations per target agent | docs/help/entity_locations.md | | Bundled cc plugin | docs/help/bundled_plugin.md | | Development guide (stack, build, tests) | docs/help/development.md | | Architecture decisions (authoritative) | docs/00_ADR.md | | Product scope | docs/01_PRD.md | | CLI surface reference | docs/04_DESIGN.md |

License

Apache 2.0