Administer

Skill library

On this page

What the library holds

Each organization has one library of skills. Members publish their own, and Valet supplies built-in and catalog skills beside them. A skill is a service, like a website, an app or an agent: it shares one set of names with them, has an author and a star, and every change to it creates a new release. To write and publish a skill, see Skills.

Built-in

What it is
Valet's governance skills, such as governance, security, and dependencies. Attach one without a copy to hold Valet's default.

Catalog

What it is
Skills Valet curates, such as deep-research and weekly-report. Install one as a copy, or attach it as catalog:<name>.

Your organization

What it is
Skills your members write or adapt. Attach one as org:<name>.

The agent's folder

What it is
A skills/<name>/ directory committed with one agent. It's already part of that agent.

Built-in skills answer coding agents until your organization publishes its own skill under the same name. They aren't services, so a hosted agent that attaches one holds Valet's default, and moves to your organization's copy once you publish it. No website, app or agent can take a built-in's name. See Replace a built-in skill.

A skill in an agent's own folder ships with every release of that agent. You change it only by changing the agent's source.

valet skills catalog lists the catalog.

How skills reach agents

Coding agents discover skills through Valet's MCP server or the CLI with the valet skill. No attachment is needed. Hosted agents carry only the skills attached to them. See Governance for how agents choose which skills to read.

How changes reach agents

Coding agents get the latest release on their next read unless they request an earlier one. Hosted agents receive updates through new agent releases, as described below.

Releases

Every publish that changes a skill's content creates a new release, numbered v1, v2, and so on. Publishing identical content creates none. Releases never change once they exist. The skill page and list_releases list them, and get_skill with release reads any one. To go back to an earlier release, read it and publish its files again, which makes a new release.

A hosted agent always carries the skill's latest release. There are no pins: attaching org:data-storage@v2 attaches the latest release, and Valet says it ignored the pin. When someone publishes a new release, Valet builds a new agent release for each agent that holds the skill, after a few minutes' delay. The agent uses it the next time it starts; Valet doesn't restart a running agent to deliver it. Renaming a skill reaches agents the same way. To keep a hosted agent's instructions fixed, commit them to the agent's own skills folder.

Attachments

Members attach skills with the attach_skill MCP tool, valet skills attach, or the Skills section of an agent's Blueprint in the dashboard. Attaching or detaching a skill rebuilds a running agent and redeploys it so the change takes effect. Skills attach to agents only; artifacts and apps don't carry skills.

Valet enforces these rules:

  • Names are unique per agent. An attached skill can't share a name with another attached skill or with a skill in the agent's folder.
  • Deploys check attached skills, not hints. A deploy stops when an attached skill has no release or shares a name with a skill in the agent's folder. A skill that valet.yaml lists but the agent doesn't hold shows in the Blueprint as a suggestion and blocks nothing.
  • Agents can't rewrite attached skills. An agent that edits its own instructions can change the skills in its folder, but not the ones attached from the library. Those belong to the library and come back unchanged with each release.
  • An attached skill can't be destroyed. Valet refuses to destroy a skill while a hosted agent has it attached, and names the agents.

Where skill content lives

Each organization skill is a Git repository in Valet's own storage, like an artifact, app or agent. Its main branch holds the latest release, and each release is a commit and a list of the commit's files. A deploy never fetches a skill from an outside host, so a deleted repository can't break a release. In a dedicated installation, the repositories stay inside it.

A catalog skill is a template. Installing one copies its files into a skill your organization owns, which then changes only when you publish.

Manage the library

List skills

MCP tool
list_skills, list_services
CLI
valet skills --org <org>
Dashboard
Skills and Governance

Read one

MCP tool
get_skill
CLI
valet skills show <name>
Dashboard
The skill's page

See which agents use it

MCP tool
get_service
CLI
valet skills info <name>
Dashboard
The skill's page

See its releases

MCP tool
list_releases
CLI
valet skills show --release
Dashboard
The skill's page

Create one

MCP tool
create_service with kind: skill
CLI
valet skills create
Dashboard
Add skill, Write a custom skill

Publish a release

MCP tool
publish_skill with files or push
CLI
valet skills publish
Dashboard
none

Change some files

MCP tool
publish_skill with edits
CLI
none
Dashboard
Edit on the skill's page

Rename one

MCP tool
rename_service
CLI
valet skills rename
Dashboard
none

Clone the latest release

MCP tool
get_skill with git
CLI
valet skills clone
Dashboard
none

Destroy one

MCP tool
destroy_service
CLI
valet skills destroy
Dashboard
The skill's page

The Standards tab of the dashboard's Governance page shows the skills that govern coding agents, and the Skills page shows the rest of the library. The launcher lists skills too. To write a standard for coding agents to follow, see Write your own standards.

Ownership and history

Valet records who created a skill and who published its latest release. The skill's page and list_services show that author. Each release is a Git commit that records when it was saved. Skills created before author tracking show no author until someone publishes them again. Valet has no audit log of skill changes. See Organizations.

Next steps

  • Skills: write and publish a skill.
  • Governance: manage the standards coding agents follow.