Skip to main content
skilder

product

AI agent glossary: role, skill, MCP server and the terms around them

Plain definitions of the terms that come up when a company governs AI agents: role, blueprint, skill, SKILL.md, MCP, MCP server, connector, registry, context lake, execution trace and more.

Author: Nicolas Corod
  • #glossary
  • #roles
  • #skills
  • #mcp
  • #ai-agents
  • #governance
skilder mascot next to the article title on a paper background

TL;DR

A skill packages know-how, an MCP server exposes tools an agent can call, and a role bundles skills, scoped tools, context and permissions for the work one seat on the org chart does. The rest of the vocabulary describes where those pieces live and how a company controls who uses them, and what they did.

The terms below come up as soon as an organisation moves from individual AI use to shared, governed capabilities. Definitions are kept short enough to quote. Where a term has its own article, the definition links to it.

The building blocks

Agent

The AI client that reasons and acts on a request: Claude, ChatGPT, Copilot, an assistant in a code editor or an in-house agent. The agent does the thinking. What it knows how to do, and what it can reach, can be supplied from outside it.

Skill

A packaged capability that tells an agent how to do a specific task: instructions, optional scripts and reference material. Skills follow the open Agent Skills format, so a skill written once works across compatible agents. For the anatomy of a skill and how an agent activates one, see why skills make agents smarter.

SKILL.md

The file at the heart of a skill. At its simplest, a skill is a single SKILL.md file with a name, a description and instructions. The format is open and portable, but a bare file is unmanaged: it says nothing about which version is live, who may run it or what it did.

MCP (Model Context Protocol)

An open protocol for exposing tools and context to agents through a standard interface. Any client that speaks MCP can use any server that speaks it, which solves the “every agent needs its own integration with every system” problem.

MCP server

A service that exposes tools and resources over MCP: read a customer record, create a ticket, query a database. An MCP server gives an agent reach into a system. It does not tell the agent when or why to use that reach.

Tool (or connector)

A single capability an MCP server exposes, such as “search invoices” or “update an order”, pointing at an ERP, a CRM, a knowledge base or an internal API. Tools are called on demand by the agent, with no copy of the underlying data. The difference between tools and skills, and why the two compose, is the subject of composing skills over MCP.

The role and how it is made

Role

The central primitive: a role bundles skills, scoped MCP tools, context and permissions, and exposes them to any agent through an MCP server. It matches the work one seat on the org chart actually does, a claims handler, a buyer, a support lead. The role is the boundary for deployment, sharing and governance: one role, many agents.

Blueprint

The starting point for a role. A blueprint describes what a role for a given job typically bundles, and is then configured to a team’s own playbooks, systems and way of working. Roles are not shipped finished and are never ready-made: the configuration on real processes is what makes them work.

On-demand loading

Pulling skills and tools into the agent’s context at the moment a task needs them, instead of preloading them. Connecting raw MCP servers loads every tool definition into the context window before any work begins. Loading on demand keeps the context small and relevant.

Where capabilities live

Registry

The place where skills, roles and MCP servers are recorded as first-class entries, each versioned and queryable, with explicit dependencies. An enterprise MCP server registry is internal and curated: it holds only servers that have been reviewed, credentialed and scoped for the organisation. It is not a public directory of every server that exists.

Context graph

A way of organising a catalogue of capabilities as a typed graph rather than a folder of files. Each skill and tool is an entity that can be queried, versioned, signed, audited and discovered, with its dependencies explicit, so it can be reused, deprecated or updated across the organisation without copying files.

Context lake

The layer that holds what an agent needs to act, as opposed to the facts a data lake stores: business meaning, the tools an agent may use and the rules that govern it. A data lake answers “what happened”; a context lake answers “what should I do”. The full argument is in why your data lake isn’t enough for AI agents.

Plugin

An installable bundle of skills and connectors for a given assistant. Plugins solve installation. They do not solve management: versioning, permissions, observability and audit still have to come from somewhere.

skills.sh

An open directory of AI agent skills maintained by Vercel Labs. Skills that follow the open format are compatible with the skills.sh ecosystem, which is one of the sources a company can draw from.

Governance terms

Shadow AI

The uncontrolled use of public AI tools by employees on company work and data, outside any policy or audit. It is usually a sign that the approved path is missing or weaker than the personal one. See what shadow AI is and why banning it fails.

RBAC (role-based access control)

Granting access by role rather than person by person. Applied to AI capabilities, the role becomes the permission boundary: a role can only reach the tools and data it was granted.

Versioning and promotion

Every skill, tool and role carries a version. Promoting a version makes it live for everyone who uses it; rolling back returns to the previous one. Both apply across the organisation, without anyone copying files.

Evals

Evaluations run against a capability to measure how well it does its job, compared across versions. They turn “promote this version” into a decision backed by a measurement rather than a guess.

Execution trace

The record of what ran: which capability, for which person, on which task, with what result. Every task is logged and attributable to the person who ran it. An execution trace is what lets an organisation answer an auditor, scope an incident, or see which capabilities are actually used.

Using this glossary

These definitions are the vocabulary skilder works with: roles, configured from blueprints, served to Claude, ChatGPT or Copilot from one registry, with an execution trace of what ran. To see how your own organisation scores on capturing, controlling and tracing its AI capabilities, the 10-question AI check-up takes two minutes and asks for no email.

Frequently asked questions

What is an AI role?

An AI role bundles the skills, scoped MCP tools, context and permissions that match the work one seat on the org chart actually does, and exposes them to any MCP-compatible agent through an MCP server. It is the unit that gets deployed, shared and governed. A role starts from a blueprint and is configured on the organisation's own processes.

What is the difference between a skill and an MCP server?

A skill carries know-how: instructions, scripts and reference material that tell an agent how to do a task, packaged at its simplest as a single SKILL.md file. An MCP server carries reach: it exposes tools the agent can call in another system, such as a CRM or an ERP. A skill says how; an MCP server provides the means.

What is a blueprint in AI roles?

A blueprint is the starting point for a role. It describes the skills and connectors a role for a given job typically needs, and is then configured to a team's own playbooks, systems and way of working. Roles are not shipped finished: the configuration on real processes is what makes them useful.

What does on-demand loading mean for AI agents?

On-demand loading means skills and tools are pulled into the agent's context at the moment a task needs them, instead of being preloaded. Connecting raw MCP servers loads every tool definition into the context window before any work begins; loading on demand keeps the context small and relevant.

Related articles