Skip to content
Quilium Docs
Back to site
Reference

MCP skill library

The library the MCP server serves to agents on demand — kinds, taxonomy, composition rules, audiences, and how a skill is discovered. Nothing here is installed or managed by hand.

Two different things are called "skills" around Quilium, and only one of them is your problem.

What it is Who handles it
The skills you install Two packages, quilium-builder and quilium-webmaster, that teach an agent the process around the CMS — build order, the CLI loop, the Liquid layer. You, once per project: npx skills add quiliumio/quilium-skills. See Prompting well for Quilium, section 2.
The MCP skill library — this page Around fifty knowledge modules served live by the MCP server, each encoding one canonical CMS pattern. The agent, by itself, through get-skills and get-skill.

The second kind is meant to be invisible. You do not install it, choose from it, update it or keep it in sync: it is served from the server, so it is always current, and the agent pulls what the task needs. If you are simply using Quilium with an agent, you can stop reading here — the practical page is Prompting well for Quilium.

What follows is internal structure. It is documented for people building on top of the library — writing skills, auditing what a given profile can reach, debugging why an agent loaded one guide and not another. None of it is a step to perform. For the rationale, see AI, MCP and skills.

The five skill kinds

The kind is the L1 type. It is encoded both as a CMS category and as the first slug segment.

Kind Purpose
meta How the library itself works.
reference Syntax and catalogues.
practices How to do a thing well.
tasks An atomic capability.
processes Multi-step orchestration.

Three-level taxonomy

get-skills returns a three-level tree. The slug encodes the path: <type>-<domain>-<skill>.

Level What it is Slug segment Example
L1 type segment 1 reference
L2 domain (sub-category) segment 2 reference-fields-catalog
L3 skill (the leaf) segment 3 the markdown guide

The L2 domain is omitted when a type carries fewer than three skills — for example meta-getting-started has no domain segment.

Composition

Skills reference one another through related:. The dependency graph is acyclic and flows in one direction only:

A skill of kind may reference
processes tasks, practices, reference
tasks practices, reference
practices reference
reference nothing

Cycles are forbidden.

Audience model

Each skill is scoped to an audience. There are two:

Audience CMS role Sees
webmaster editor meta, reference, practices, and webmaster tasks. Never tasks-types-*, tasks-config-*, tasks-templates-*, or architecture processes.
webdesigner owner, admin, webdesigner The full library.

The scope is enforced by a triple lock: the CMS profile field, MCP ?mode=webmaster URL routing, and the skill's own audience field.

Discovery

The agent calls get-skills once to retrieve the taxonomy, then loads each relevant skill with get-skill <slug>. Both are ordinary MCP tools; nothing else triggers a load.

Walk both branches. Skills hang off the tree in two places. The META group carries skills directly under groups[*].skills[*]; most groups attach them through sub-categories under groups[*].categories[*].skills[*]. Read BOTH or you miss skills.

The live taxonomy returns five groups: META (1 skill), BEST PRACTICES (5 categories), PROCESSES (3 categories), REFERENCE (9 categories), and TASKS (11 categories) — roughly 50 skills. Skills marked status: experimental are skipped unless asked for; deprecated skills are treated as nonexistent.

HTTP fallback. The library is also served at /<locale>/md, /<locale>/md/<slug>, /<locale>/skills-manifest, and /<locale>/skills-feed. Prefer get-skill over raw HTTP.

The golden rule

Never duplicate. When tempted to copy a YAML block, a snippet, or a checklist into a second skill, add a related: reference to the source instead. An agent follows related: transitively, so the content stays reachable. Duplication causes drift.

Related


Last updated September 6, 2026.