geelen&company
a project · built with agentic engineering

Ring20 Skills
Platform

An agent is only as safe as the tools you hand it. This platform turns operational capability into a governed catalogue of skills — each one discoverable, permissioned, and callable three ways — so an agent can act on real infrastructure without going off the rails.

Status In active use · 24 skills, 132 operations
Repository Internal build — source not public
Stack Python · MCP · Qdrant
Ring20 Skills Platforma CLI and an MCP server, not a GUI — the interface is the tool catalogue itself

Screenshot — pending

Purpose & goals

The Ring20 Skills Platform manages a library of skills — small, focused modules that wrap a homelab or infrastructure API in a standard, safe interface — and exposes them three ways: as command-line tools, as tools on a Model Context Protocol server, and as results from a semantic search. The goal is to give an agent a discoverable, permissioned catalogue of things it is allowed to do, so operational work is capability-gated rather than improvised. (Its working name carries the number “150” as a target; the platform today runs a real, working set of skills.)

Modules & architecture

Access flows through three layers — the git repository of skills, a semantic search index over them, and the protocol server that presents them to agents. Around that sit the moving parts: the per-skill modules; the MCP server with role-based filtering; the sk command-line tools; role definitions; and a generated manifest that turns each skill into callable tools. Governance is not an afterthought — a four-layer authority model (role, maturity, escalation, sovereignty) plus a credential gate with one-time-password two-factor authentication decides what may actually run. The platform is itself an agentic-engineering project, with a single dedicated agent as its only mutator.

Functionality & technology

The catalogue currently holds about two dozen skills exposing over a hundred operations, covering virtualization (Proxmox), DNS and DHCP, reverse proxying, ad-blocking, object and backup storage, CI/CD, secrets management, home automation, knowledge and task search, alerting, certificates, and deployment workflows. The stack is Python and shell, with a manifest generator turning skills into tools, Qdrant for semantic discovery, the Model Context Protocol over both streaming and standard transports, a secrets manager for credentials, and hosted models for embeddings. There is no end-user GUI by design — the platform is the interface an agent reaches through.

← Back to Agentic Engineering

VALUES INTENT AND OUTCOME OVER METHOD AND PRECEDENT