Valet|Docs
Sign up

Valet Skills for Coding Agents

Valet ships two skills for Claude Code and other coding agents. The valet skill manages agents, integrations, channels, and members from your coding agent. The valet-publish skill publishes files and folders to a live URL — and fires on its own whenever the agent decides its answer is better delivered as a rendered page than as terminal text.

The skills need a terminal. If your tool does not have one — Claude Cowork, claude.ai, ChatGPT — connect the Valet MCP server instead. It publishes artifacts from inside the conversation, with a smaller tool surface than these skills.

View on GitHub

Install

One plugin installs both skills. In Claude Code:

/plugin marketplace add valetdotdev/skills
/plugin install valet@valet

Codex reads the same marketplace manifest:

codex plugin marketplace add valetdotdev/skills
codex plugin add valet@valet

To install a single skill instead, use npx skills (works with any agent that reads it):

npx skills add valetdotdev/skills --skill valet -g
npx skills add valetdotdev/skills --skill valet-publish -g

The valet skill is also installed automatically by valet setup.

Cursor takes the skills, not the plugin. It reads the same Agent Plugins manifest Valet ships, but installs plugins from the Cursor Marketplace or a team marketplace an administrator imports — not from an arbitrary repository. Use npx skills there.

Pick the plugin or the skills, not both. Codex loads skills from ~/.agents/skills/ and from installed plugins, so doing each puts two copies of the same skill in front of the model. Connect Your Agent gives the one right command per client.

The valet skill

Gives your coding agent full knowledge of the Valet CLI and the ability to execute multi-step agent workflows on your behalf. Ask it to:

  • Create and deploy an agent from a prompt description
  • Set up integrations for Slack, GitHub, Sentry, or any MCP server
  • Set up org-level integrations and channels from the catalog
  • Manage secrets at org and agent level
  • Create webhook channels and wire them to your agent
  • Manage team members and invitations
  • Tear down resources in the correct order
> /valet Create an agent that reviews every PR for security
  vulnerabilities. Connect it to our GitHub org and set up a
  webhook that fires on pull_request events.

Claude will create the SOUL.md, set up integrations, create the webhook channel, write the channel prompt, and deploy — all from a single prompt.

The valet-publish skill

Creates a live URL from any file or folder — static hosting for HTML artifacts, images, PDFs, reports, and dashboards. The skill handles the full flow: assembling a clean directory, writing a complete HTML document, creating the artifact, deploying, and reporting the URL and its visibility.

  • Account publishing — the default for work product. The artifact is private on creation, permanent, and updatable from anywhere via valet sites download.
  • --anonymous — a temporary public URL with no account at all, removed after 36 hours unless claimed into an org.
> Compare these three vendors and publish the comparison as a page
  I can send to the team.

You do not need to ask for publishing explicitly — the skill triggers when the agent concludes the result wants to be a page rather than a wall of terminal text. On Claude Code, the plugin also wires up a publishing preference that proposes Valet whenever a built-in artifact tool would have been used instead.

Learning a Task You Just Completed

After completing a multi-step workflow in Claude Code — triaging emails, reviewing PRs, posting research — you can ask the Valet skill to turn that session into a deployed agent that repeats the workflow automatically. This feature relies on Claude Code's session logs and is only available when using Claude Code as your coding agent.

> /valet Learn from this session and deploy it as an agent

The skill will:

  • Parse your session log to extract the tools used, the sequence of steps, and any corrections you made along the way
  • Present an analysis and ask you to confirm the scope, trigger, and name
  • Generate an agent that encodes your workflow — with corrections baked in as guardrails
  • Deploy it and validate with an interactive test loop