Changelog

What changed in skillsvault, newest first. The CLI tells each machine when a newer release is out; re-run the install script to upgrade.

30 October 2026

Skills as prompts in claude.ai, Claude Desktop and Claude Code. Every approved skill is now an MCP prompt: pick it from the chat's + menu or as a slash command. Blocked skills aren't offered, and each use is audited under your name. MCP docs
Agents can search the registry. list_skills takes a query ("fill pdf forms") and returns the best matches first, each with its description.
A health endpoint at /api/health for your uptime monitor.
Evaluation status on every skill. Its page shows the last trigger test (how many prompts routed as expected, precision and recall) and says when the description has changed since.
The scanner flags manipulative instructions. Text that tells the agent to keep something from the user ("do not mention this to the user", "silently send…") or to ignore its rules is flagged high, and calls to another MCP server's tools are named. These checks come from Anthropic's enterprise vetting checklist.

Sign in from any MCP client. claude.ai, ChatGPT, Codex, Cursor and other clients that support MCP sign-in connect with just https://skillsvault.io/api/mcp: you sign in, pick your organisation, and every skill loaded is audited under your name. Disconnect apps from your account page. How it works
Policy exceptions. A block links to a request. A reviewer sees what the exception would change, what the skill does and how often it was blocked, then grants it for 7, 30 or 90 days. Every machine honours it until it lapses, and the requester is reminded two days before. Exceptions
Skill owners. Every skill has an accountable owner. The dashboard flags skills without one, and the evidence report names each owner.
Similar skills. A skill's page, and the review of a new one, name skills in the registry that do the same job.
Registry search matches words in any order, best match first.
⌘K jumps to any page or skill.
Who ran it. The audit log, alerts, dashboard and skill pages name the person behind each run.
A Claude Code plugin and the MCP Registry. claude plugin marketplace add skillsvault/skillsvault-skill; listed as io.skillsvault/skillsvault.
CLI 0.6.7 to 0.6.9. The gate honours exceptions and links a block to a request; a machine gated through the Claude Code plugin reports as gated.
Public Explore. Search skills.sh and check any public skill for free, then add it to your registry.

Admin activity. Who changed the rules (policies, skills, members, keys, alerts), recorded by the database for every path.
Weekly digest for reviewers: what the gate did, and what's waiting on a person.
Policy preview on real runs. Before you save a rule, see what it would have done to the last 30 days of runs.
Imported skills are flagged when their source changes or disappears.
A public security page.

Members propose skills; governance approves. A member's skill or change, from the editor or skillsvault publish, waits for review, with an AI review ready before anyone decides.
Alerts for blocks, warnings and submissions in Slack, Teams, Discord, or as signed JSON for your SIEM.
Capability policies. Block a skill for what it does (runs shell commands, installs packages, reaches a host), enforced on the machine from CLI 0.6.4.
A printable evidence report for your auditors.
skillsvault pull verifies every skill's content against the signed policy bundle (CLI 0.6.3).
Accessibility. Every console page meets WCAG 2.1 AA in both themes.

The editor tests your skill. Trigger tests show which prompts would invoke it, and a scorecard explains what to fix.
AI review on every publish, with a deterministic scanner underneath it.
Rate limits and security headers on every endpoint.