ToolHive

ToolHive is Stacklok's answer to the mess of running Model Context Protocol servers: instead of each server sitting on your machine with your machine's credentials and network access, ToolHive drops every server into its own container with a minimal permission profile, network filtering, and secret management. It can even containerize a server whose authors never shipped an image, pulling it straight from a package manager.

It ships in three editions — a desktop UI for individuals, a Go CLI (thv) for developers, and a Kubernetes operator for platform teams — plus a Registry Server implementing the official MCP Registry API for curating trusted catalogs, and a Gateway (Virtual MCP Server) that collapses backends behind one endpoint with OIDC/OAuth SSO, OpenTelemetry traces, and Prometheus metrics. Everything is Apache-2.0 and self-hostable.

The honest caveat: this is infrastructure, not magic — you still have to point it at an identity source and write the permission policies, otherwise you get a sandbox without governance. The browser-based cloud UI was retired, so plan around the desktop app and CLI. For teams standardizing MCP in production, it is the most complete open-source control plane available.

1. Install the CLI (thv) or the desktop app; Docker, Podman, or Colima required for the runtime.
2. Discover a server from the registry and run thv run <server> — it launches containerized with a default permission profile.
3. Tighten the permission file: network access, secrets, and tool filtering per server.
4. Connect your AI client (Claude Desktop, Cursor, VS Code); for teams, deploy the Kubernetes operator and Registry Server.