Skills
Understand the core Milkey tool model and how skills move through your application.
Skill Model
A Milkey skill is authoritative, reusable guidance that an AI system can retrieve at runtime. Instead of placing a large prompt template directly into every app or project, Milkey keeps the skill in hosted infrastructure and delivers it only when your workflow asks for it.
This lets teams keep the app stack in control while centralizing skill discovery, retrieval, and maintenance.
Canonical Tools and Provider Aliases
Milkey exposes three canonical tools. Provider adapters may register underscore-safe aliases, but the Milkey backend still uses the canonical names.
resolve-skill -> milkey_resolve_skillget-skill -> milkey_get_skillget-skill-reference -> milkey_get_skill_reference- resolve-skill: find the best skill for a query, optionally scoped by category.
- get-skill: fetch the full content for the chosen skill.
- get-skill-reference: fetch one specific reference document by full reference slug.
Why the alias mapping exists
Reference Documents
Some skills expose references alongside the main skill body. References are optional supporting documents that can be fetched independently when the workflow needs more detail.
The reference slug format is always skill-slug/reference-slug. Invalid formats should be treated as application errors rather than retryable transport failures.
Delivery Model
Milkey delivers the same skill model across both public access paths:
- MCP: the agent calls Milkey tools through the hosted MCP server.
- SDK: your provider adapter registers Milkey tools and your app executes the loop.
Whether you access Milkey through hosted MCP or the SDK, the product story remains the same: Milkey resolves and retrieves skills; your application or agent decides what to do next.