How to run Claude Code in a Mac terminal — and five of them in parallel

One Claude Code session is productive. Five Claude Code sessions aren't productive times five. They're chaos times five, unless you've solved the window-management problem underneath.

This guide walks through three things, in order: installing Claude Code on Mac, running it in a terminal, and the actual hard problem, which is keeping track of what a bunch of parallel agents are doing at once.

Installing Claude Code on Mac

Two paths. Pick whichever fits your setup.

Via npm (most common):

npm install -g @anthropic-ai/claude-code

You need Node 18 or later. Homebrew users can get it with brew install node first. Once installed, claude --version should print a version string.

Via the native installer: download from code.claude.com. This gives you the same CLI, pre-signed, no Node required.

First launch will prompt you to authenticate. Claude Code opens your browser, you sign in with the same Anthropic account that backs Claude.ai, and the token lands in your macOS keychain. No separate API key needed for a subscribed account.

Verify with:

cd ~/code/some-repo
claude

You should see the Claude Code prompt, ready for a task.

Running one Claude Code session

The minimum-viable workflow is boring on purpose:

1. cd into your project directory (a git repo gives Claude better context, but it isn't required).
2. Type claude.
3. Give it a task in plain English: "add JWT auth middleware to the user router".
4. Watch it read files, propose changes, write them, and (if you let it) run your tests.
5. Approve, reject, or redirect as it works.

This is fine for one task. The whole thing starts to be useful when you stop doing one task at a time.

Running five Claude Code sessions in parallel

The pattern I actually use: one terminal per concurrent task. Different features, different repos, different branches, all running their own Claude Code session at the same time.

It looks like this, conceptually:

• Terminal 1: api repo, feat/jwt-auth branch — Claude adding auth middleware.
• Terminal 2: api repo, feat/tests branch (git worktree) — Claude writing integration tests.
• Terminal 3: frontend repo, feat/dashboard — Claude building the new dashboard page.
• Terminal 4: frontend repo, main — Claude reviewing an open PR.
• Terminal 5: infra repo — Claude writing a Terraform module.

Each session is an independent process. There's no shared state by default. A Claude Code session in Terminal 1 doesn't know what Terminal 3 is doing. If two sessions are touching the same repo, use a git worktree per session to keep their filesystem views isolated.

That's the easy part. The hard part is the windows.

The window-management problem

Start with one terminal. Easy. Open a second, alt-tab between them, still fine. Open a third, maybe tile them manually. At five terminal windows open on a single screen, you're doing one of the following:

• Losing track of which agent is waiting for your input.
• Alt-tabbing through windows like a Wordle game trying to find the one that finished five minutes ago.
• Manually arranging the windows in a grid every time you add or close one.
• Switching to the other monitor hoping three of them are over there.

None of these scale. There are three honest ways out.

Option 1: tmux or Zellij

The classic Unix answer. tmux (or the more modern Zellij) gives you panes inside a single terminal, keyboard-driven navigation, persistent sessions that survive network drops.

If you already use tmux and it fits your head, run Claude Code inside tmux panes. It works fine, Claude is just a CLI. No integration needed.

If you don't already use tmux, I wouldn't pick it up just for this. There's a real configuration and muscle-memory cost (prefix keys, session names, pane arithmetic) and you're now managing two layers of keyboard shortcuts: tmux's for windows, Claude's for code. Learning tmux because you want to run multiple AI agents is like buying a kiln because you needed a coffee. See tmux alternatives on Mac for the longer version of this argument.

Option 2: Multiple terminal windows, manually arranged

If you're disciplined and your laptop has a big screen, this works. Command-N for a new Terminal.app or iTerm2 window, drag into position, live with the overlap. Everyone does this at some stage.

It stops scaling at about four windows, when arranging them evenly becomes a second job. You can buy one of the dozens of tiling window managers for macOS (Rectangle, Magnet, Moom, Amethyst), which helps with the arranging but still requires you to be the one pressing the keys each time a new terminal opens.

Option 3: A terminal that auto-tiles

This is the category Waffle sits in, and in fairness, the reason I built it. Open a terminal, it's fullscreen. Open a second, they split 50/50. Four become a 2×2 grid. Nine become 3×3. You don't arrange anything.

The workflow for running five Claude Code sessions in parallel becomes:

1. Press ⌘N for a new terminal. It tiles automatically.
2. Type claude to start a session.
3. Repeat four more times. You now have a 3×2 grid of live Claude Code sessions, all visible at once.
4. Use ⌘←/⌘→/⌘↑/⌘↓ to move focus between agents.
5. Use ⌘⏎ to zoom the current one to fullscreen while it works; ⌘⏎ again to return to the grid.

Waffle also auto-detects git repos and colour-codes sessions by project. A grid of three repos × two agents each shows you at a glance which agent is in which codebase, because each project gets its own accent colour and its own filter tab.

Keyboard flow reference

The shortcuts I actually use running five agents in parallel. Works in Waffle; substitutes in tmux or iTerm2 exist but are longer.

Practical tips for running many agents at once

Use git worktrees for same-repo sessions

If two Claude Code sessions need to work in the same repository, give each one its own worktree so they don't fight over the working tree. git worktree add ../repo-feature feat/x then cd into the worktree in the new terminal. This lets Claude run tests and commit without one session stomping another.

Name your task out loud in the first prompt

When you're running five agents, it's easy to lose track of what each one is working on. The first prompt you give each Claude Code session is effectively its label in the grid. Write it like a Git branch name: short, concrete, unambiguous.

Don't ask too many agents to run tests at once

Five agents running tests in parallel will start fighting over the same ports, test databases, and cache directories. Stagger test runs or keep one session as a "reviewer" that runs tests after others finish writing code.

Keep one terminal free for you

Not every terminal in your grid needs a Claude. I usually keep one in the corner for git status, gh pr list, shell commands, a REPL. Anywhere I want to poke around without bothering an agent.