How to run multiple Claude Code or Codex sessions on Windows

The short answer: open one terminal per session, give each session its own working copy of the code (a git worktree is the cleanest way), and keep each one on a task that does not touch the others' files. That works, today, with nothing but Windows Terminal and git. It also stops working somewhere around the third session, when you can no longer tell which agent is working and which one has been silently waiting on you for twenty minutes. This guide covers the manual setup honestly, where it breaks, and the purpose-built way to do it.

Why run more than one session at all

Claude Code and Codex both work the same way: you give the agent a task, and it spends the next several minutes reading code, editing files, and running commands without needing you. Those minutes are the whole opportunity. While one agent fixes a bug, a second can write tests and a third can update documentation. You are not typing faster - you are running more work in parallel. One developer supervising three or four sessions ships noticeably more than one developer driving a single session and watching it think.

Method 1: multiple terminals, one session each

The zero-setup version. Open Windows Terminal, open a tab per session, start claude or codex in each, and give each one a task. Two habits make this survivable:

• Name the tabs. Right-click a tab, choose Rename Tab, and label it by task ("auth bug", "tests", "docs"). Untitled tabs all look identical the moment you need to find the one that stalled.
• Split panes for a shared view. Alt+Shift+D splits the current pane, so you can see two or three sessions side by side without switching tabs. Past three panes the text becomes too small to read.

The catch: if every session is in the same folder, the agents will edit the same files. Two agents rewriting the same file at the same time produces the kind of mess that erases everything parallelism saved you. Which is why you also want method 2.

Method 2: git worktrees, so sessions cannot collide

A git worktree gives you a second full checkout of your repository, on its own branch, in its own folder - sharing the same underlying repo. Each agent works in its own worktree, so nothing collides and each branch can be reviewed and merged on its own:

cd C:\repos\myapp
git worktree add ..\myapp-auth feature/auth
git worktree add ..\myapp-tests feature/tests

Now start one session in C:\repos\myapp-auth and another in C:\repos\myapp-tests. When a branch is merged, remove its worktree with git worktree remove ..\myapp-auth. Terminals plus worktrees is the honest manual answer, and plenty of developers run exactly this.

Where the manual setup breaks

Every problem with the terminal approach is a supervision problem, not a git problem:

• Nothing tells you when an agent needs you. Claude Code and Codex stop and ask questions. In a background tab, a question looks exactly like progress: silence. The expensive failure mode is an agent that has been idle for half an hour while you assumed it was working.
• You become the status board. With four tabs open, "how is everything going?" means clicking through four tabs and reading scrollback in each. You do that every few minutes, all day.
• The sessions are chained to that window. Close the laptop, lose the desk, or just walk to a meeting, and you are blind until you are back in front of those exact terminals.

Method 3: an orchestrator built for this

DevThrottle is an open source Windows app built specifically for running multiple command-line coding agents at once. It does not replace Claude Code or Codex - it runs the same CLIs you already use and adds the supervision layer the terminals are missing. Every session appears as a card on one board showing Working, Waiting on you, or Done, so the blocked agent is impossible to miss. A background service keeps sessions alive when no window is open, and you can check in and answer questions from any browser or your phone - by voice if you want. Everything runs on your own machine; your code never leaves your hardware.

For the general habits that make a fleet of agents productive - task splitting, batching reviews, isolation - see how to run multiple AI coding agents without losing track.

Start with two sessions today

If you are new to this, do not start with six agents. Add one worktree, run two sessions on genuinely separate tasks, and feel where your attention goes. When you catch yourself cycling through tabs asking "is anyone stuck?", that is the moment the manual setup has hit its ceiling. Create a free account, download the Windows app, and you will have your first supervised session on the board in about ten minutes.