Skip to main content

Discover and install skills in your AI client, and run vMCP locally

This week, the ToolHive Desktop UI installs skills directly into the AI clients you already use, and Virtual MCP Server (vMCP) is now supported locally. The Registry Server also surfaces version information in list views.

Discover, install, and build skills in the Desktop UI​

Skills are reusable instruction bundles that teach AI clients how to perform specific tasks. ToolHive Desktop UI v0.33.0 ships a dedicated Skills section with three tabs (Registry, Local Builds, and Installed) that cover the full skill lifecycle:

  • Browse and install from the registry. Search the configured catalog, open a skill's detail page to read its rendered SKILL.md, and install it straight into Claude Code, Cursor, or any other client ToolHive manages. Choose User scope to install globally for your user, or Project scope to install into a specific Git repository.
  • 130 skills available out of the box. The bundled registry ships with skills ready to install across common workflows.
  • Connect your own skills registry. Point ToolHive at a registry your team maintains to surface your organization's curated skills alongside the bundled ones. See Use the registry for how to add a custom registry in the desktop app.
  • Build your own skills locally. The new Local Builds tab packages a folder on your machine into an OCI artifact and installs it into a client for testing. When it's ready to share, publish it to a Registry Server.
  • Track and uninstall in one place. The Installed tab shows every skill ToolHive has added to your clients, with scope and project-root indicators, plus a confirmation flow for removing them.

Run vMCP locally via the CLI​

You no longer need a Kubernetes cluster to try vMCP. With the new local CLI mode, you can take the MCP servers you already run on your machine and present them to your AI client as a single endpoint, with the same aggregation, tool routing, and optimizer that vMCP provides in production.

This is a good fit when you want to:

  • Evaluate vMCP locally before rolling it out centrally via Kubernetes.
  • Develop locally against several MCP servers behind one endpoint instead of configuring each client one server at a time.
  • Cut down the tool list your AI client sees with the built-in optimizer, so prompts stay fast and focused.

Start with the local vMCP CLI guide. When you're ready for a shared production deployment, the same concepts carry over to the Kubernetes VirtualMCPServer resource.

Version metadata in registry list responses​

If you build a UI on top of the Registry Server, or use a client that does, catalog views now load faster and ship with richer per-entry context. The ListRegistryEntries endpoint embeds version metadata inline with each entry in the list response. Previously, clients had to follow up with a per-entry detail request to retrieve version information, turning every list view into an N+1 fetch pattern. With version fields included in the list payload, UIs and registry clients can render staleness signals, "update available" badges, and version-aware filters directly from a single paginated response. This applies to both MCP server entries and skill entries served through the registry-scoped browse endpoints.

Getting started​

For detailed release notes, check the project repositories:

You can find all ToolHive documentation on the Stacklok documentation site.