The Claude stack glossary

The Anthropic product family has gotten dense. There's Claude.ai, Claude Desktop, Claude Cowork, Claude Code, plus Skills, Projects, Scheduled Tasks, Routines, /loop, MCP, the Agent SDK, and Managed Agents. Half of those names sound interchangeable. They are not. This is the cheat sheet that keeps you (and your customers) on the right one.

Last verified against primary Anthropic docs on May 13, 2026. If something below is wrong, tell me and I'll fix it.

The four surfaces

There are four top-level places to run Claude. They all share your account login but they are not the same product, and a Skill you upload to one does not automatically show up in another.

Claude.ai

The default. A chat product on the web, iOS, and Android. Includes Projects (chat workspaces with custom instructions and files), Artifacts, Research, web search, Memory, and connectors. The model you talk to by default is Sonnet 4.6.

This is the surface you use from your phone. It is also where most small business owners spend ninety percent of their Claude time.

Claude Desktop

The Mac and Windows native app. Looks like Claude.ai with extra hosting hooks. Drag a file in. Drop in a screenshot. Add local MCP servers via claude_desktop_config.json.

The reason Desktop matters: Cowork only runs inside Claude Desktop. Not on the web. Not on the phone. If you want Claude doing real autonomous work for you, you need the desktop app installed on a Mac or Windows machine that is awake when the work runs.

Claude Cowork

Anthropic's product for non-coding agentic work. Lives as a mode inside the Claude Desktop app, not as a separate app. You switch into Cowork from a tab. It does file org, data extraction from screenshots and receipts, document drafting, light browser automation, Slack and Chrome integration, computer use as a fallback.

Cowork is the SMB-friendly side of the agentic stack. The dev-shop side is Claude Code.

Available on Pro, Max, Team, Enterprise. Not on Free.

Claude Code

The CLI and IDE tool that started as a coding assistant and is now the broader agent runtime. Lives in your terminal, on claude.ai/code on the web, and as IDE plugins (VS Code, JetBrains).

Despite the name, Claude Code is not just for software developers. It's the right tool any time you want Claude to read your local filesystem, run scripts, schedule cloud-hosted automations (those are Routines, see below), or do an agentic task that doesn't fit the Cowork box.

Included in Pro ($20). You do not pay extra for Claude Code on top of your Claude subscription. This is the single most-missed fact in the entire pricing page.

Features that live across surfaces

Skills (also called Agent Skills)

A markdown file with frontmatter (name, description) plus optional supporting code and reference files. Claude loads the Skill on demand when a task matches the description. You can pin a Skill to a Project so it's always loaded in that workspace.

Skills are account-level: upload a Skill once and it is available in Claude.ai, Cowork, Claude Code, and Office add-ins without re-uploading. You can also pin a Skill to a specific Project to make sure it always loads in that workspace.

Anthropic ships pre-built Skills for pptx, xlsx, docx, and pdf document handling. You can also write your own, like the Scrappy Recipes on this site.

Projects

A claude.ai container that bundles custom instructions, knowledge files, and conversation history into one persistent workspace. Per-conversation context.

Projects and Skills solve different problems. A Project is a workspace with your stuff in it. A Skill is a reusable instruction package the model loads when relevant. You can pin Skills to Projects.

Custom Instructions

Plain free-text personalization. Lives inside a Project, or at the account level. The lightest possible Claude customization.

The mental model: Custom Instructions is a sticky note. A Project is a desk with that sticky note plus your files plus your past conversations. A Skill is a reusable manual you can put on any desk.

MCP (Model Context Protocol)

An open spec for plugging AI models into outside tools and data. Gmail. Google Calendar. Your file system. Custom apps. Slack. Anything.

Supported on Claude.ai, Claude Desktop, Claude Code, and Cowork. The plain-English version: MCP is the connector layer that turns Claude from "a model that answers questions" into "an assistant that can actually touch your stuff." For small business owners, the practical entry point is Google Calendar and Gmail connectors.

The scheduling muddle (the most confusing area, get this right)

Three names. Three surfaces. Not interchangeable.

| Name | Lives in | What it is | |---|---|---| | Scheduled Tasks | Cowork (inside Claude Desktop) | Recurring tasks that run on your Mac or PC. Requires the machine to be awake and Claude Desktop to be open at trigger time. Frequency: hourly, daily, weekly, weekdays. | | Routines | Claude Code (cloud-hosted) | Cloud-run automations. Trigger by schedule (≥1 hour cadence), API call, or webhook. Runs on Anthropic infrastructure. Your machine can be off. Per-day limits: 5 on Pro, 15 on Max, 25 on Team and Enterprise. | | `/loop` | Claude Code (a session command) | A polling command inside a single Claude Code session. /loop 5m check the deploy reruns the prompt every five minutes for the life of the session. Auto-expires after 7 days. Session-scoped, not persistent. |

Do not say "Routine" when you mean a Cowork Scheduled Task. They are different products. The shorthand for "a thing on a clock" is "Scheduled Task" if you're in Cowork (most SMB scenarios) and "Routine" if you're in Claude Code (more technical, cloud-hosted).

There is also a fourth thing called Desktop Scheduled Tasks that lives inside Claude Code as a local-runner variant. Runs on your machine on schedule but persists across restarts, unlike /loop. This one is a power-user feature; most people don't need to know it exists.

Models, briefly

| Model | What it's good for | |---|---| | Claude Opus 4.7 | The flagship. Hardest reasoning, most expensive, default model on Enterprise. Available on Pro and up with the right model selector. | | Claude Opus 4.6 | Previous flagship. Still available, still excellent, slightly cheaper. | | Claude Sonnet 4.6 | The everyday workhorse. The default model on Claude.ai and Cowork. What you're talking to when you open the app and don't change anything. Released Feb 17, 2026. | | Claude Haiku | The smallest and fastest. For high-volume cheap-and-fast tasks. Not what you want for nuanced writing. |

The 1M context window is available on Opus 4.6, Opus 4.7, and Sonnet 4.6 via the API and Claude Code cloud sessions. On Claude.ai chat the practical context cap is 500K for all paid plans, including Max.

Practical rule for SMB owners: stay on Sonnet 4.6 (the default) for everything until you hit a wall. If a long document keeps getting cut off or a hard reasoning task feels weak, switch to Opus.

Pricing, May 2026

| Plan | Price | What you get | |---|---|---| | Free | $0 | Basic Claude.ai chat, pre-built Skills (no custom upload), no Cowork, no Claude Code. | | Pro | $20/mo ($17/mo annual) | Full Claude.ai, Cowork, Claude Code, custom Skills. Routines: 5/day. The cheapest way to access the full toolkit. | | Max 5x | $100/mo | Everything in Pro plus higher rate limits, Opus 4.7 at 1M context, Routines: 15/day. The "I rely on Claude every day" tier. | | Max 20x | $200/mo | Same toolkit as Max 5x, higher rate limits again. For very heavy users. | | Team | $25/seat/mo ($20 annual). Premium seats $100/seat/mo annual ($125/seat/mo monthly). | Org features, admin controls, shared Skills via Skills directory. | | Enterprise | Custom, 50-seat minimum | Everything plus SSO, audit logs, default Opus 4.7. |

The fact most SMB owners miss: Pro at $20 already includes Cowork and Claude Code. You don't need to upgrade for "the real tools." You need to learn what's already in your subscription.

Common confusions, with the truth

| You might think... | The truth | |---|---| | Cowork is a separate app | It's a mode inside the Claude Desktop app for Mac or Windows. Not on web or mobile. | | Claude Code is just for developers | No. It's any agentic workflow that needs filesystem access, scripted automation, or cloud Routines. SMB owners with one terminal open can use it. | | Skills sync across my Claude account | They do. Skills are account-level and available in Claude.ai, Cowork, Claude Code, and Office add-ins without re-uploading. | | Routine and Scheduled Task are interchangeable | They are not. Scheduled Task is Cowork. Routine is Claude Code cloud. | | I need Max to get Cowork | No. Pro ($20) includes Cowork and Claude Code. | | Claude Code SDK is the developer tool | It was renamed. The current name is Claude Agent SDK. | | Anthropic Console / Claude Console | The developer console is now at platform.claude.com, not console.anthropic.com. |

Which surface for which job (SMB version)

| The work | Surface | Why | |---|---|---| | Triaging customer emails, drafting replies, captions, light writing | Claude.ai (web or mobile) | Lightest weight. Phone-accessible. No setup. | | Reading a contract or screenshot, anything with files | Claude Desktop | Native drag-drop, plus local file access. | | "Run this every Monday morning while I sleep" on your real files | Cowork Scheduled Task | Local-machine agent on a clock. Easiest to set up. | | "Run this every weekday at 6am from the cloud, machine can be off" | Claude Code Routine | Cloud-hosted, doesn't need your machine. Slightly more setup. | | One-off agentic job, e.g., "go through these 200 invoices and pull totals" | Cowork (handed a task) | Cowork's bread and butter. | | Complex workflow with shell + filesystem + scheduled cloud + GitHub | Claude Code | Has every knob. | | "I'd rather not figure this out solo" | A Scrappy Hour with me | One hour, on Google Meet, $0 first or $150. |

Things new since January 2026

If you read this with old context in your head, here's what changed.

- Sonnet 4.6 (Feb 17, 2026) replaced Sonnet 4.5 as the default Claude.ai model. - Opus 4.7 (April 16, 2026) is the new flagship; Opus 4.6 still ships. - 1M context went GA across Opus 4.6 / 4.7 / Sonnet 4.6. - Routines in Claude Code launched April 14, 2026. Cloud-hosted scheduled automations. Different product from Cowork Scheduled Tasks. - Claude Code SDK was renamed Claude Agent SDK. - Managed Agents beta launched (hosted REST API for production agents). - Anthropic Labs launched as an experimental product incubator. First public release: Claude Design. - Cowork enterprise admin controls launched April 9, 2026. - Help center moved from support.anthropic.com to support.claude.com. The product blog is at claude.com/blog; anthropic.com/news continues operating for company news. Both are active.

Receipts

Primary sources for everything above, with last-updated dates where Anthropic showed them.

- Claude pricing - Pro plan, what's included (≈ Apr 2026) - Max plan, both tiers (April 7, 2026) - Cowork product page - Get started with Cowork (≈ late April 2026) - Schedule recurring tasks in Cowork (April 9, 2026) - Introducing Routines in Claude Code (April 14, 2026) - Run prompts on a schedule (Claude Code docs) - Introducing Agent Skills (Oct 16, 2025; updated Dec 18, 2025) - Agent Skills overview - Use Skills in Claude - Claude Opus 4.7 announcement (April 16, 2026) - Claude Sonnet 4.6 announcement (Feb 17, 2026) - 1M context GA - Context windows on paid plans (≈ May 12, 2026) - Agent SDK overview - Managed Agents overview - Model Context Protocol - Anthropic Labs intro

Where this came from

I help small business owners use Claude every week, and I kept getting the names wrong myself. This is the doc I now pin to every Claude project that writes about Anthropic products. It's CC0. Install it as a Skill in your Claude account, paste it into a Project as custom instructions, or share it with the bookkeeper who keeps mixing up Routines and Scheduled Tasks.

If a name shifts or a price changes, this page updates. The version number in the frontmatter goes up. The receipts above stay current.

Jamie at Scrappy Start