A package registry for the things your AI agents depend on

What it does

It’s a central registry — written in Go, shipped as a Docker Compose stack with a bundled PostgreSQL — where teams publish and discover the building blocks of agentic systems: MCP servers (from npm, PyPI, OCI images, or remote endpoints), skills (SKILL.md bundles with docs and code), agents (an identity plus its server and skill dependencies), and reusable prompts. A CLI (arctl) and a web UI handle scaffolding, publishing, versioning, and running, locally or on Kubernetes via Helm.

The interesting bit

The boring part is the value: it treats AI artifacts like ordinary packages, with versions, metadata, and environment-variable declarations, so “which MCP server can I actually trust” becomes a curation question instead of a Slack archaeology dig. It also auto-generates IDE config for Claude Desktop, Cursor, Kiro, and VS Code, and pairs with agentgateway so a newly deployed MCP server is picked up by clients without reconfiguration.

Key highlights

• One catalog for MCP servers, agents, skills, and prompts, with versioning and governance controls
• arctl covers the full lifecycle: scaffold, build, push, register, run
• Deploys locally or to Kubernetes from the same registry; Helm support included
• Auto-configures popular AI clients and wires routing into agentgateway, an AI-native reverse proxy with JWT auth and OTEL observability
• Web UI at localhost:12121 for browsing and managing artifacts without the CLI

Caveats

• The gateway integration is really a pitch for a sibling project; agentregistry alone doesn’t handle auth or traffic — it’s the catalog and delivery layer
• Community infrastructure is young: no community meetings yet, and the roadmap lives in GitHub issues
• Requires Docker Desktop with Compose v2+ for the quick-start path

Verdict

Worth a look if your team’s MCP servers live in a spreadsheet and every onboarding is a bespoke Docker ritual. Solo developers with two servers and a prayer can skip it for now.