Skills Library
Agent Skills are packaged instructions, scripts and resources that AI coding tools and assistants load on demand. The Skills Library is where you curate, scan and distribute Skills: find the Skills already in use on your endpoints, import Skills from public GitHub repositories or upload your own, review what people submit, install approved Skills for an audience, and publish them to AI coding tools.
The page has five tabs: Installed, Discovered, Library, Sources and Publishing. Each tab has its own address, so a link or a reload opens the same tab. Add repository and Upload Skill sit in the page header on every tab.
Summary tiles
Skill details
Select a Skill's name on any tab to open its details: overview, security scan, capability evidence, declared tools and dependencies, compatibility requirements, instructions, file manifest and validation diagnostics. Read the instructions and the file manifest before you install or approve a Skill.
Security scan
Every Skill is scanned in the background after it is added. The scan combines automated pattern checks with an AI review, and each finding shows its severity, category, source (AI review or Pattern check) and the file and line it came from.
Skills that contain scripts are tagged with a script count, so you can see them before you review or install.
A scan, like provenance and license metadata, is evidence to weigh, not a guarantee that a Skill is safe.
Installed
Skills installed for your organization. Each row shows the Skill's source and version, audience, security scan, compression, distribution channels (Self Service and Claude Code), declared capabilities, risk and criticality. A Required in Claude Code tag marks Skills that install automatically.
- Settings opens Configure Skill, where you change the audience, the requested compression and whether the Skill is required in Claude Code.
- Uninstall records the installation as uninstalled. History and the pinned package remain.
Discovered
Skills the Endpoint Agent has seen on devices, whether or not you installed them. For each Skill you see the applications using it, devices and users, scope, risk, criticality, what the agent reports for it (approval and execution state) and when it was last seen. Use this tab to find unvetted Skills already in use.
To block or unblock a discovered Skill, select Manage in Inventory. Skill rules are managed in one place, in Inventory and the Endpoint Agent Policy Engine.
Library
Skills available from your sources. Each entry shows provenance, license, package tokens, compatibility, declared capabilities, security scan and state. Review a Skill here, then select Install. Selecting a repository on Sources opens the Library filtered to that repository's Skills.
Sources
The repositories and uploads the Library is built from. Each source shows its kind, pinned commit, import summary (discovered, conformant, compatible, quarantined, imported and failed), license and trust, and refresh state.
Pinning a source to a commit means the Library does not change silently when the upstream repository changes. Refresh deliberately, and review what changed.
Add Skills to the Library
Add a public GitHub repository
- Select Add repository in the page header.
- Enter the Repository URL, for example
https://github.com/owner/repository. Only public github.com repositories are accepted. - Select Add and refresh. The repository is imported and pinned to an immutable commit, and you land on Sources.
Refreshes run only when you start them, and never follow other hosts.
Upload a Skill
- Select Upload Skill in the page header.
- Drop or choose a Skill package: a
.zipwithSKILL.mdat its root or in one folder, or a singleSKILL.md. Up to 8 MB. - Tick I confirm this organization has the right to distribute this Skill to its users.
- Select Upload.
Uploading the same Skill name again adds a new version. Earlier versions stay installable, and nothing reaches users until you install it for an audience.
Review a submitted Skill
People can propose Skills from Self Service with Submit a Skill. A submission waits on Sources with a Review button, and a banner shows how many are waiting. Submitted Skills stay out of the Library until approved.
- On Sources, select Review on the submission.
- Select Open Skill details and read its instructions and files. Skills that include scripts deserve a closer look.
- Optionally add a Note to the submitter.
- Select Approve or Reject.
Approving adds the Skill to the Library. It reaches people only after you install it for an audience. The submitter sees the outcome under Your submissions in Self Service.
Install a Skill
- On Library, find the Skill and review its provenance, license, declared capabilities and security scan. Open the Skill's details for scan findings.
- Select Install.
- In the Install dialog, set:
- Audience: All users (tenant-wide) or Smart Group. For a Smart Group, search for and pick one of your Smart Groups.
- Requested compression: Text compression · pending or Original package.
- Required in Claude Code: install the Skill automatically for everyone in the audience instead of offering it.
- If the license is not MIT, Apache or BSD, tick I attest that this tenant has distribution rights. Uploaded Skills already carry the uploader's confirmation.
- If the Skill declares shell/Bash or wildcard tool preapproval, read the declarations under Administrator capability review and tick I reviewed these exact declarations for this immutable version. QuilrAI records the version and a digest of the declarations as review evidence. This does not grant Endpoint permissions.
- If the scan rated the version critical, enter a Reason for overriding the scan. It is recorded with the installation.
- Select Install for audience. The Skill appears on Installed.
To change the audience, compression or the Claude Code requirement later, use Settings on the Installed tab.
Self Service
Installed Skills are available to people in Self Service under Skills:
- Available to you lists the Skills shared with everyone or with their teams. People can copy a Claude Code install command, or download a Skill to unzip into their Skills folder or upload as a custom Skill in Claude's settings.
- Your submissions lists the Skills they submitted and each one's state: Awaiting review, Approved, Not approved or Withdrawn by an admin.
Publishing
Publishing sends the Skills installed in QuilrAI to AI coding tools, so they arrive on people's computers and stay up to date. Pick the tool in AI coding tool:
Both Claude tools read the same organization Skills feed with one organization token. Generating it is a one-time setup. The token is shown only once, so copy it or download the deployment files right away.
Manage token (under Step 1 · Access) has two actions:
- Rotate token: the current token stops working immediately, for Claude Code and Claude desktop alike. Redeploy the new token so every device keeps receiving updates.
- Revoke token: both tools stop receiving Skill updates immediately. Skills already installed on computers stay until they are removed there.
Skills installed for a Smart Group reach people only when the deployed file or helper includes each person's work email. Turn on Include each person's work email in the deploy step. The page warns you when Smart Group Skills would not reach anyone with the option you picked.
Claude Code
- Open Publishing, pick Claude Code and select Set up Claude Code, then Generate.
- In Step 2 · Deploy to developer machines, pick where you manage developer machines and follow the steps:
The scripts add a drop-in file that merges with any Claude Code managed settings you already deploy. If you already ship a Claude Code profile on macOS, add the keys to it instead of installing a second profile. Smart Group Skills need Intune · Windows or Jamf · macOS with Include each person's work email turned on; the other targets deliver only Skills for all users.
- Developers receive Skills automatically after deployment. QuilrAI cannot see which devices received the file.
Claude desktop
Claude desktop must be set up for third-party inference. A Claude desktop signed in with a claude.ai account does not read an organization marketplace.
- Open Publishing, pick Claude desktop and select Set up Claude desktop, then Generate. If you already deployed a token for Claude Code, paste it in Token already deployed for Claude Code? instead. It stays in your browser and fills the installers.
- Step 2 · Install the QuilrAI helper on each computer. Claude desktop runs this small helper to send the token when it fetches Skills. Pick Jamf · macOS, Intune · macOS or Intune · Windows and download the installer. To try it on one computer first, run it in Terminal (
sudo sh quilr-skills-claude-desktop-macos.sh) or in PowerShell as administrator. Turn on Include each person's work email for Smart Group Skills; where the signed-in account isn't the work email, deploy the email script or profile too. - Step 3 · Add the QuilrAI marketplace. Either in the app (Help → Troubleshooting → Enable Developer Mode, then Developer → Configure Third-Party Inference → Plugins → Add marketplace → Marketplace URL, enter the values shown, then Apply Changes or Export a profile for every computer), or add the
allowedPluginMarketplacesvalue shown to the Claude desktop device profile you already deploy. - Step 4 · People install their Skills. Restart Claude desktop on each computer. Skills appear under Plugins → Organization and update on their own.
If the Skills don't appear:
For some organizations whose Claude desktop already uses the QuilrAI LLM Gateway for inference, the page also offers QuilrAI LLM gateway sign-in (no token). Claude desktop then reuses its gateway sign-in, and no token or helper is installed on computers: you only add the marketplace and restart. Contact your QuilrAI representative to enable it.