Tal EckerlingAI Specialistעברית
Building with Claude CodeLesson 15 of 16 · 6 min

Subagents, MCP, automations and Claude Code in the cloud

In this lesson you'll learn

  • Subagents
  • MCP connections and plugins and hooks
  • Working in parallel in the cloud and on a schedule
  • Managing costs

Listen to this lesson

8 min · 7 parts · AI narration

First: the right dose

This lesson covers Claude Code's advanced abilities. But I'm opening with a dosage warning, because I've seen this movie: you discover this world, connect ten tools, build a whole setup of agents, and two weeks later you can't remember what does what.

My rule: add an advanced ability only when something specific hurts, and that ability solves it. Go through the lesson, mark what's relevant today, and come back when the need shows up. And every ability here draws on your allowance, so the right dose is good for your wallet too.

Subagents: a team instead of a single employee

A subagent is a helper Claude sends off on a task, with its own clear desk. It investigates, summarizes, and brings back only the conclusion to the main conversation. Remember the cluttered desk from lesson 13? Here the messy investigating happens at the subagent's desk, and your conversation stays clean.

You don't need to set anything up. Just ask:

Ready prompt

"Use a subagent to check how all the pages on the site are built, and come back to me with just the list of pages and the structure of each one."

Ready prompt

"Before we change anything: send a subagent to go through all the content files and come back with a list of every place the old price appears. Just a list, don't change anything."

Ready prompt

"We've finished the new page. Have a fresh subagent go through it as if it were a client seeing it for the first time: what's unclear, what's missing, and what breaks the design."

The third one is my favorite: fresh eyes catch what the eyes that did the work miss, and the official docs also recommend a reviewer with a clean context. Just remember that every subagent uses its own tokens from the same allowance. For simple tasks, you can ask for it to run on Haiku, the economical model.

Connections and extensions: MCP, plugins and hooks

MCP (Model Context Protocol) is an open standard that connects Claude to outside systems. You add a connection with one command in the terminal, for example to Notion:

claude mcp add --transport http notion https://mcp.notion.com/mcp

Inside Claude Code, the /mcp command shows your connections and handles signing in. You'll find vetted connections in Anthropic's directory, and connectors you've already connected in your Claude account show up in Claude Code on their own, when you're signed in with the same account.

What changed: I used to recommend five or six connections at most, because every connection took up space in the context. As of October 2026, tool search is on by default: at the start of a conversation only the tool names load, and a tool loads in full only when it's needed. So an extra connection barely weighs on the context. My advice today: connect what you really use, turn off what you don't in /mcp, and choose only servers you trust. The official docs warn that a server that brings in content from outside exposes you to prompt injection, and every connection gives Claude real access to a system. The limit today is trust and oversight, not space.

Plugins pack skills, subagents, hooks and connections together, in one install. You manage them with /plugin, and a plugin you installed in your Claude account comes to Claude Code on its own. Install only from a source where you know who's behind it.

Hooks are rules that are always enforced. An instruction in CLAUDE.md is a recommendation Claude does its best to follow. A hook is an action that runs automatically every time something specific happens, for example blocking every edit in a sensitive folder. You don't write them yourself, you ask: "Set up a hook that blocks any change in the originals folder." The /hooks command shows what's set up.

In parallel, in the background, in the cloud and on a schedule

Once you're used to one Claude, you discover the next stage: several jobs at once.

  • Simplest: another terminal window, another conversation. Works when the tasks don't touch the same files.
  • In the desktop app: the New button in the sidebar opens a parallel conversation.
  • In the background: /background sends the conversation to run in the background and gives you your terminal back. The claude agents command shows on one screen what's running, what's waiting for you and what's finished (this view is still in an early stage).
  • In the cloud: at claude.ai/code the task runs on Anthropic's servers and keeps going even when you close your computer. For that you need a GitHub account and a code repository.

My advice: start with two jobs in parallel. Your role is to supervise, and past a certain point, the supervising becomes the bottleneck.

Script mode. The claude -p command followed by a request runs Claude Code without a conversation, so you can plug it into automated processes. A run like that starts in manual mode today, and actions that need approval are declined, unless you allowed them in advance. The official advice: try it on two or three files before you run it on a hundred.

Scheduled tasks. The /schedule command creates a scheduled task in the cloud, which runs even when your computer is off. It's still at an early stage, and to work on a code repository it needs access to GitHub. Need files from your computer? A scheduled task in the desktop app runs with access to your files, but only when the app is open and the computer is awake.

And research: /deep-research followed by a question sends out searches in many directions, cross-checks sources and brings back a report with sources. It's an example of a dynamic workflow, where Claude orchestrates many subagents in parallel, and it uses your allowance to match. On the Pro plan, you turn the ability on first through /config.

What it costs

The allowance is shared between Claude in the app and Claude Code. The levers, according to the cost management page:

  • Tracking: /usage shows how much you've used and what used it. On paid plans you'll also find a breakdown and recommendations there, and the f key applies whatever can be applied.
  • Model: the default is Opus 5.5. The docs note that Sonnet handles most coding tasks well and costs less, and that Haiku suits simple subagents. Switch with /model.
  • Tidiness: /clear between tasks. A shorter context means less usage.
  • Past the cap: /usage-credits opens the usage credits settings, which keep you working at API prices.
  • API credits: according to the pricing page, the Max plan now includes $100 to $200 a month in API credits, and the Team plan up to $500 a month for the whole team together, for building on Claude Platform.

Try it now

  1. Open Claude Code in the folder of a project that already has a few files.
  2. Paste:
Ready prompt

"Send a subagent to go through all the files in the project and come back with the three most important problems it finds: broken text, a link that doesn't work, or inconsistent design. Just a list, don't change anything."

  1. When the list comes back, pick one problem and ask Claude to fix it.
  2. Type /usage and see how much this work used.
  3. Type /mcp, check what's connected, and turn off what you don't use.

What to take with you

Subagents keep the conversation clean, MCP connects systems, hooks enforce rules, and the cloud works even when your computer is closed. Add an ability only when something hurts that it solves.

Want to keep your progress?

Sign up once, at no cost: the lessons you finish are marked in your library, and the shelf of recipes and tools opens up.

Join the academy
Subagents, MCP, automations and Claude Code in the cloud | The Claude Academy | Tal Eckerling