

The control plane for agent skills.
One skill that connects your agent to your organisation's registry —
install what's approved, block what's banned, audit everything.
Works with Claude Code · Codex · Pi · every other harness
⚡ Add it to your agent
npx skills add skillsvault/skillsvault-skill
That's it. Your agent reads the skill and bootstraps itself — it installs the
skillsvault CLI, asks you to sign in once, then pulls your organisation's
approved skills into every harness on the machine:
curl -fsSL https://skillsvault.io/install.sh | sh
skillsvault login
skillsvault pull --all
As a Claude Code plugin
claude plugin marketplace add skillsvault/skillsvault-skill
claude plugin install skillsvault@skillsvault
The first time Claude uses it, run /mcp and sign in to skillsvault: you pick your organisation, and no key is pasted or stored in a file. The plugin adds:
- The skillsvault skill: the same skill as above.
- The skillsvault MCP server:
list_skills, load_skill and get_skill_file let Claude read your organisation's approved skills. All three are read-only. A skill your policies block is refused, with the reason.
- The gate hook: before Claude runs a skill, subagent or typed
/command, it asks the skillsvault CLI whether your organisation allows it. A block shows the policy's reason and a link to request an exception. The hook does nothing until the CLI is installed, and steps aside if skillsvault install-hook already wired the gate.
In other agents
The same repository is a plugin for other agents too. Each one signs you in to skillsvault the first time it connects:
- Cursor, Codex and ChatGPT: add the MCP server
https://skillsvault.io/api/mcp. Marketplace listings are on their way; the plugin already follows the Agent Plugins standard (plugin.json, mcp.json, skills/).
- Gemini CLI:
gemini extensions install https://github.com/skillsvault/skillsvault-skill
- Anything else that speaks MCP: add
https://skillsvault.io/api/mcp. It is listed in the official MCP Registry as io.skillsvault/skillsvault.
What it sends, and where
- MCP server. It calls
https://skillsvault.io/api/mcp as you, once you sign in. It sends the skill names Claude asks for, and receives your organisation's skill content. Each load is recorded in your organisation's audit log.
- Gate hook. It runs on your machine against a signed copy of your organisation's policies, with no network call per decision. The CLI reports each decision (skill name and version, decision, reason, machine name) to your organisation's audit log. It also sends a periodic heartbeat (CLI version, whether the gate hook is wired, and the names of the skills installed on the machine), so your organisation can see which machines are covered.
- Nothing else. Nothing is sent anywhere other than skillsvault.io. Your prompts and code are never sent. See the privacy policy.
🎬 See it work
🧭 Why
Agents pick up skills from everywhere — repos, marketplaces, teammates, other
agents. Nobody knows which version runs where, and a poisoned or outdated
skill looks exactly like a good one. skillsvault gives your organisation one
source of truth:
| |
|---|
| 📦 Versioned registry | Every skill content-addressed, versioned, and signed. pull installs the approved version — and removes banned ones. |
| 🚦 The gate | Every invocation checked locally, before it runs: allow / warn / halt. Works offline from a cached policy bundle. |
| 🧾 Audit trail | Who ran which skill version, on which host, under which policy — append-only and exportable. |
| 🕸️ Every harness | One policy for Claude Code, Codex, Pi, and anything else that runs skills. |
🛠️ What your agent does with it
Once connected, the skill teaches your agent to govern itself:
skillsvault whoami
skillsvault install-hook
skillsvault pull --all
skillsvault gate --skill <org/skill>
skillsvault publish <path>
skillsvault status
Three rules the skill enforces, always:
- If the gate halts a skill → the agent does not run it. Ever.
- The approved version from
pull beats any sideloaded copy.
- A halt is never bypassed — the verdict names the approved replacement.
🔎 Scan your skills in CI
Publishing skills? Check them on every push and pull request. The action installs the CLI (verified against the release checksums), scans every folder with a SKILL.md, and fails the job when one pipes a download into a shell, reads secret paths or hides content. Nothing is sent anywhere.
name: skills
on: [push, pull_request]
jobs:
scan:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: skillsvault/skillsvault-skill@main
with:
path: skills
fail-on-flag: true
To see findings as code-scanning alerts on the pull request, write SARIF and upload it:
permissions:
security-events: write
steps:
- uses: actions/checkout@v4
- uses: skillsvault/skillsvault-skill@main
with:
path: skills
sarif: skillsvault.sarif
- uses: github/codeql-action/upload-sarif@v3
if: always()
with:
sarif_file: skillsvault.sarif
Locally, the same check is skillsvault scan --dir skills --fail-on-flag (add --sarif for SARIF).
📚 Learn more
License
Apache-2.0 — see LICENSE. This covers the skill and plugin in this repository; the skillsvault service is governed by its terms.