Claude Glossary of Terms

Anthropic ships new features faster than they ship clear names for them. By the time you’ve figured out what an “artifact” is, someone’s already talking about a “live artifact” and a “subagent” and you’re not sure if those are three things or one. This glossary is the cheat sheet I wish I’d had a year ago.

Artifact

What it is. A self-contained piece of work—a document, a chart, a snippet of code, a one-page HTML report—that Claude generates in a side panel during your conversation. It’s static. It doesn’t refresh. It doesn’t pull live data. It’s a finished thing you can copy, save, or share.

Example. You ask Claude to draft a one-page summary of your quarterly numbers. It builds the page in an artifact panel. You read it, tweak it, save it.

When to reach for it. Any time you want a clean deliverable that lives outside the chat window. Memos. Drafts. Code you’re going to paste somewhere else. One-shot reports.

Connector

What it is. A live link between Claude and one of your outside tools—Gmail, Slack, Google Calendar, Notion, Box, Asana, and so on. Once you install a connector and grant permission, Claude can read your data there and, where you allow it, take action.

Example. You install the Gmail connector. Now you can ask Claude, “What unread emails do I have from clients this week?” and it actually checks. (You can also create custom connectors for other apps, but that’s a conversation for another time.)

When to reach for it. When you want Claude to work on real data from your real tools instead of stuff you paste into a chat. The first three connectors most people install are Gmail, Calendar, and Slack.

Cowork (Cowork mode)

What it is. Anthropic’s desktop app for non-developers. Think of it as Claude with a body—file access to your Mac, the ability to control your computer, scheduled tasks that run while you sleep, plugins you can install, and live artifacts that update themselves.

Example. You open Cowork in the morning. It greets you with a briefing it generated at 6:50 a.m. while you were getting the kids ready for school.

When to reach for it. When you want Claude to do work on your computer, on a schedule, with access to your files and apps. Cowork is currently in research preview, so things move fast.

Dispatch

What it is. When Claude hands a focused subtask off to a subagent—a fresh, isolated Claude session with its own context and tools—to run in the background and return a single summary. The main conversation keeps moving. The dispatched work happens in parallel.

Example. You ask Claude to research five companies. Instead of doing them one at a time and filling your context window with five long research reports, it dispatches five subagents at once and gives you back a clean comparison.

When to reach for it. Whenever a task has clear chunks that don’t depend on each other. Parallel research. Code review across multiple files. Surveying a lot of documents quickly. Trust Claude to dispatch when it’s the right move; you rarely have to ask for it by name.

Knowledge

What it is. Files you attach to a Project as permanent background context—PDFs, Word docs, Markdown notes, transcripts, anything. Claude reads them in every conversation inside that Project. Knowledge is project-scoped: a file you load into your Newsletter Project doesn’t follow you into your Client Strategy Project.

Example. You drop your brand guide and three back issues of your newsletter into a “Writing” Project’s Knowledge. Every chat in that Project now starts with Claude already knowing your voice.

When to reach for it. When the same reference material needs to be in front of Claude across many conversations in a single domain. Brand guides. Pricing sheets. Style guides. Recurring client briefs.

Live Artifact

What it is. A self-contained HTML page Claude builds for you that persists across sessions and refreshes its data from your connectors every time you open it. Unlike a regular Artifact, it’s not a snapshot. It’s a living dashboard you can come back to.

Example. You ask Claude to build a “tasks waiting on me” page. It saves as a Live Artifact. Tomorrow you click it, and the list rebuilds itself from the current state of your Asana, your Reminders, and your inbox.

When to reach for it. Anything you’d otherwise want to ask Claude to regenerate every morning. Status pages. Personal dashboards. Trackers. Anything that pulls from a connector and needs to stay current.

MCP (Model Context Protocol)

What it is. The open protocol Anthropic published in late 2024 that defines how Claude talks to outside tools. An “MCP server” is a small program that exposes a tool, an app, or a data source in a way Claude can use. Connectors are MCP servers that Anthropic, or a third party, has wrapped up and made one-click installable.

Example. When you install the Notion connector, what you’re really doing is installing the Notion MCP server. The “connector” label is the friendly name; “MCP” is the plumbing.

When to reach for it. You don’t, usually. You reach for the connector. But knowing the word matters because Claude, Anthropic, and the developer community use it constantly, and you’ll see it in product release notes and plugin descriptions. When Claude doesn’t provide an official link, you can create a custom connector if the appo vendor provides MCP or API access.

Memory

What it is. Persistent notes Claude keeps about you—your role, your preferences, your recurring projects, things you’ve corrected it on—that travel across every future conversation. Memory lives in a private folder on Anthropic’s side, not inside any single chat.

Example. You tell Claude once, “I prefer short emails. Never use exclamation points.” Six weeks later, in a completely new chat, it remembers.

When to reach for it. You don’t reach for it; Claude writes to memory on its own when something feels worth keeping. What you can do is ask it directly to remember something, or to forget something. Memory is global. Knowledge is project-scoped. Don’t confuse the two.

Plugin

What it is. An installable bundle that packages Skills, Connectors, slash commands, and other capabilities together into one piece you can drop into Claude Code or Cowork with a single click. Plugins are how a team or a vendor distributes a complete workflow instead of asking you to wire up six separate things.

Example. A “Hot Leads Report” plugin might bundle the weekly research skill, the CRM connector, a few prebuilt slash commands, and a scheduled task that fires every Monday morning. You give it to your client, and they install with one click.

When to reach for it. When you want all the pieces of a workflow installed together, or when you want to share a workflow with your team or a client without making them assemble it from parts.

Plugin Marketplace

What it is. The catalog where you browse and install Plugins—yours, Anthropic’s, and the wider community’s. Marketplaces live inside Claude Code and Cowork, and you can subscribe to multiple marketplaces from different sources.

Example. Anthropic maintains an official marketplace. Your company can publish its own marketplace and put internal plugins in it. The team installs from the company marketplace; everyone gets the same tools, the same way.

When to reach for it. When you’re looking for a ready-made workflow before you build one from scratch. Most of the time, somebody’s already done it.

Project

What it is. A workspace inside Claude that bundles a system prompt, a set of Knowledge files, a custom configuration, and a persistent chat history all scoped to one purpose. Think of a Project as Claude, configured for this one thing I do over and over.

Example. A “Newsletter” Project holds your style guide and brand voice as Knowledge, has a system prompt that reminds Claude how you write, and keeps every newsletter conversation in one tidy place.

When to reach for it. Any recurring use case. Writing. A specific client. A specific book you’re working on. The moment you find yourself pasting the same context into chat for the third time, that’s a Project. The good news is that you can retroactively add conversations to projects.

Scheduled Task

What it is. A task Claude runs automatically at a defined time—every weekday at 6 a.m., every Sunday at noon, once an hour, once tomorrow—without you in the loop. The task is just a prompt and a cron expression (specific time); Claude does the work and notifies you when it’s done.

Example. “Every morning at 6:50, build my daily briefing, deploy it to Netlify, and send me a Slack DM with the link.”

When to reach for it. Anything that runs on a rhythm. Morning briefings. Weekly research reports. Nightly journal entries. Hourly inbox sweeps. If you’d otherwise have to remember to do it, schedule it.

Skill

What it is. A focused instruction set, usually a SKILL.md file plus a few optional helper scripts, that Claude loads when a user request matches the skill’s triggers. Skills are how you encode a procedure once and have Claude execute it consistently every time. Think of it as a SOP for AI.

Example. A “newsletter workflow” skill that tells Claude exactly how to draft, apply smart typography, save, and deploy a weekly newsletter, invoked whenever you say “let’s write this week’s Insider.”

When to reach for it. Any procedure you do the same way every time. Skills are the building blocks Plugins are made of. Single source of truth lives in one place; behavior is consistent across every conversation.

Subagent (Agent)

What it is. A separate, isolated Claude session that the main Claude session spins up to handle a focused subtask. The subagent has its own context window and its own tool access, does the work, and returns one summarized message. Its full back-and-forth doesn’t pollute the main conversation.

Example. You ask Claude to audit a fifty-page contract for risk clauses. It spawns a subagent to do the heavy reading, and you get back a clean summary of findings rather than fifty pages of running commentary.

When to reach for it. Same answer as Dispatch: you usually don’t ask by name. You give Claude a task that’s clearly multi-part or research-heavy, and trust it to delegate. Subagents are the what; dispatching is the act of using them.

Cross-references

• Artifact vs. Live Artifact. Artifacts are snapshots. Live Artifacts refresh themselves from connectors. If the data needs to stay current, you want Live.
• Memory vs. Knowledge. Memory is global and travels everywhere. Knowledge is project-scoped and only loads inside one Project. Pro Tip: If you want certain knowledge to be available globally, turn the knowledge file into a Claude skill.
• Skill vs. Plugin. A Skill is one procedure. A Plugin is a bundle of Skills plus Connectors and other pieces packaged for distribution.
• Connector vs. MCP. A Connector is the friendly name. MCP is the protocol underneath. Same thing, different vocabulary.
• Subagent vs. Dispatch. A Subagent is the helper. Dispatch is the act of sending work to one.

This glossary is maintained as Anthropic ships new features. Last reviewed: 2026-05-18.