Multi-agent installs
A single skill can be installed into several AI agents at once, and into either a project or your user-global location. This page explains the model so you can predict where content lands.
One skill, many agents
Section titled “One skill, many agents”GSKILL treats each agent (Claude Code, Codex, Cursor, Antigravity CLI, OpenCode, OpenClaw, Hermes Agent) as a target. When you install a skill, you choose which agents receive it:
- Pass
--agent <id>one or more times to target specific agents. - Pass nothing and GSKILL installs into the agents it detects in the project (by their markers:
.claude/,.codex/,.cursor/,.opencode/oropencode.json(c),.hermes/).
The lockfile records the full set of target agents and the exact path the skill was installed to for each one, so a restore reproduces the same multi-agent layout everywhere.
Shared-location agents
Section titled “Shared-location agents”Antigravity CLI, OpenCode, OpenClaw, and Hermes Agent read project skills straight from
.agents/skills/, where GSKILL commits the content. For them the committed copy is the install: the
lockfile records the target with mode shared, and nothing else is written.
The consequence is that targeting cannot hide a project skill from these four agents. A skill you
install only for claude still sits in .agents/skills/, so all four can load it. Use targeting to
control what GSKILL records and installs globally, not what a shared-location agent can see.
Why detection matters
Section titled “Why detection matters”Detection makes the common case effortless: if your project already uses Claude Code and Codex, a plain
gskill add ./skill installs into both without you listing them. But it also means a project with no
detected agents and no explicit --agent has nowhere to install — GSKILL writes nothing and exits 9
rather than guessing.
Project vs global scope
Section titled “Project vs global scope”- Project scope (default) keeps skills with the project: content committed at
.agents/skills/<name>/, agent directories linking into it. This is what you commit and reproduce per repository. - Agent-global installs (
--global) write verified copies into each agent’s own user-global location, shared across projects. They sit outside the repo-owned model, backed by the same clone cache.
The lockfile records the agents and install mode, so intent and reality stay aligned.