Skill Scope: Workspace, Agent, or Shared
A skill is a folder. Where you put that folder decides who can run it. Marketplace installs and custom skills both land somewhere; mixing those locations without an allowlist is how a personal helper shows up in a customer channel.
Exact directory names follow your installed docs. Typical layout: a per-user managed tree (often under the OpenClaw config dir), a per-agent tree, and a workspace
skills/. Confirm with docs.openclaw.ai for your version. Install mechanics: Installing Skills.Three places
| Scope | Put it there when | Do not put it there when |
|---|---|---|
| Workspace | The procedure is about this repo or this client folder only. | You want the same skill on every agent you run on the machine. |
| Per-agent | Only the mail bot or only the coder should see it. | You keep copying the folder between agents by hand. |
| Shared / managed | ClawHub installs and house-standard tools every agent may use. | The skill holds a token or a destructive shell path. Those stay per-agent plus env secrets. |
Precedence and allowlists
When the same skill name exists in more than one tree, the tighter scope should win (workspace over shared). If your build exposes an allowlist or denylist per agent, set it. Specialist teams from the 9.5 wizard need this or every specialist inherits every ClawHub install.
v2026.10.9 also confines mirrored skill sync destinations to the sandbox skills/ root and stops using frontmatter-controlled names as destination paths. Do not fight that by writing skills outside the sandbox root.