Skip to content

Glossary

Term Meaning
Agent A model running in a loop that chooses tools, observes results, and continues until the task is done.
Agent Skills standard The open, cross-vendor spec for SKILL.md folders (https://agentskills.io). Claude, Cursor, Codex, and others implement it.
AGENTS.md A plain Markdown instructions file at a repo’s root or in subfolders. It’s read by Cursor, Codex, and others, and it’s the vendor-neutral counterpart of CLAUDE.md.
Client (MCP) The component inside a host that holds one connection to one MCP server.
Context window The model’s working memory for a conversation. Everything you load (rules, skill bodies, tool schemas, file contents) competes for it.
Delegation A parent agent handing a task to a subagent. It’s driven by the subagent’s description.
Description The frontmatter text a model reads to decide whether to use a skill, subagent, or tool. It’s the single most important line you write.
Elicitation (MCP) A server asking the user for extra input partway through an operation, via the client.
Eval / eval set A fixed set of prompts with expected outcomes, used to measure whether a skill, agent, or tool behaves correctly and to catch regressions.
Frontmatter The YAML block between --- lines at the top of a Markdown file.
Hook A deterministic script run by the host on an agent event (before a tool call, after an edit, and so on).
Host (MCP) The application the user interacts with (Cursor, Claude Desktop, Claude Code) that runs MCP clients.
Inspector @modelcontextprotocol/inspector, the official interactive tester for MCP servers.
MCP The Model Context Protocol, an open JSON-RPC-based protocol for connecting AI hosts to tools and data.
MCPB MCP Bundle (.mcpb, formerly .dxt), a one-click installable package of a local MCP server for Claude Desktop.
Orchestrator An agent whose main job is to plan and delegate to other agents.
Progressive disclosure Loading information in stages. For a skill: metadata first, then the body when it’s relevant, then reference files only when they’re needed.
Prompt (MCP) A reusable, user-invoked message template that a server exposes.
Resource (MCP) Read-only data that a server exposes by URI (files, records, schemas).
Roots (MCP) Filesystem or URI boundaries a client tells a server it may operate within.
Sampling (MCP) A server asking the client’s model to generate a completion.
SDK agent An agent created and run from code (Claude Agent SDK, Cursor SDK) rather than from a chat UI.
Skill A folder containing SKILL.md (plus optional scripts, references, and assets) that packages a procedure or domain knowledge.
stdio transport MCP over a child process’s stdin/stdout. It’s used for local servers. stdout is reserved for protocol messages, so log to stderr.
Streamable HTTP The current MCP transport for remote servers: a single HTTP endpoint with optional SSE streaming. It replaces the older HTTP+SSE transport.
Subagent A child agent with its own prompt and context window. It returns a result to the parent.
Tool A function the model can call, with a name, description, and JSON Schema input. MCP servers expose tools; hosts have built-in tools too.
Tool annotations MCP hints about tool behaviour: readOnlyHint, destructiveHint, idempotentHint, openWorldHint.
Trigger The situation that should cause a skill or subagent to be used. It’s encoded in the description.