Appearance
How to Run Multiple Claude Code Sessions in Parallel
To run multiple Claude Code sessions in parallel, give each session its own checkout of the repository and a place to keep running. The usual recipe is one git worktree per task, one claude process per worktree, and something to hold the processes: terminal tabs, tmux, or a session manager such as Clusto, which keeps every session on a daemon and shows them side by side from your desktop, phone or browser. Claude Code's own subagents and agent teams cover parallel work inside a single session.
This guide goes through each option, then shows how Clusto handles it.
Option 1: several terminal windows
Open two terminals, cd into the project in each, and run claude. It takes no setup. The catch is that both sessions edit the same working tree, so one agent's half-finished change breaks the other's test run, and both may touch the same file. It also ends when you close the terminal or the laptop sleeps.
Option 2: tmux
tmux keeps sessions alive after you disconnect and splits one terminal into panes:
bash
tmux new -s agents
# run claude, then split with Ctrl+B % and run claude again
tmux attach -t agents # later, from another SSH sessionThat fixes persistence on a server. It does not fix the shared working tree, and tmux has no idea when an agent is waiting on a permission prompt. See A tmux Alternative for AI Coding Agents.
Option 3: git worktrees
A worktree is a second checkout of the same repository on its own branch. Give each session one and the edits stop colliding:
bash
git worktree add ../myapp-auth -b feature-auth
cd ../myapp-auth && claudeLast reviewed: November 20, 2026. Claude Code has this built in. Common workflows shows claude --worktree feature-auth, which starts an isolated session in a new worktree; run it with another name in a second terminal for a second session. Cleanup is covered under Worktrees.
Each worktree still needs its own dependencies installed, and if your app binds a fixed port or uses one local database, two sessions running the app at once will fight over them.
Option 4: subagents and agent teams
Claude Code can parallelize inside one session:
- Subagents run focused tasks in their own context window and report a result back to the main agent. See Subagents.
- Agent teams coordinate several Claude Code instances with a lead, a shared task list and direct messages between teammates. The agent teams docs mark them as experimental and disabled by default, enabled with
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. They list limits that matter here: one team per session, no nested teams,/resumedoes not restore in-process teammates, and split-pane mode needs tmux or iTerm2. They also warn that teams use significantly more tokens than a single session and suggest starting with 3 to 5 teammates.
These are good when one agent should split up a job. They still run inside the lead's process on one machine, so they stop when that process does.
Option 5: a session manager
A session manager runs the sessions for you and gives you one view of them. Conductor does this on macOS with a workspace per task; see Conductor Alternative for how it compares. Clusto does it on Linux and Windows machines and keeps the sessions on a daemon, covered below.
How Clusto runs parallel Claude Code sessions
Every session lives on a daemon
Install the daemon (clustod) on the machine that should do the work:
bash
curl -fsSL https://dl.clusto.app/install.sh | shThen connect the desktop app, the Android app or the web app. Each New Session with the Claude agent preset starts another Claude Code process on that node. Sessions keep running when every client is closed, and any client reattaches to them. Setup details are in Run Claude Code on a Server or VPS and Downloads and Installation.
From a shell on the node you can start sessions in separate worktrees you created yourself. The CLI streams output but does not forward keystrokes, so you type into them from the app:
bash
clusto open --dir /srv/myapp-auth --shell claude --name "auth"
clusto open --dir /srv/myapp-billing --shell claude --name "billing"
clusto lsWatch them side by side
The Terminals view splits into panes: Alt+Shift+= splits right, Alt+Shift+- splits down, Alt+Arrow moves focus and Ctrl+Shift+M maximizes one pane. A pane is only a view, so closing it leaves the session running. See Terminal Workspace.
Each session row shows what its agent is doing: a spinner while it works, a green check when the turn is done, and an amber ? when Claude waits for an answer. Go to next waiting session in the F1 palette cycles through the ones that need you. See Agent Waiting Marker.
Or as chats
A Chat runs Claude Code headless with a conversation view, approval cards and queued messages. The Chats view splits into panes with the same keys, and Ctrl+Shift+D splits the focused pane and starts a new chat in the same workspace. You can mix providers: Claude, Codex, OpenCode or OpenRouter.
Let a team split the work in worktrees
A Clusto team is a leader plus workers, each a headless Claude, Codex, OpenCode or OpenRouter agent. Workers spawned with assign or handoff get an isolated git worktree under .clusto/worktrees/<worker-id>, a private port range and a database name of their own, which covers the two conflicts worktrees alone leave open. At most five workers are active at once; extra assignments queue until a slot frees up.
bash
clusto team create --name review-team --working-dir /srv/myapp
clusto team ls
clusto team assign <TEAM_ID> --prompt "Write tests for the billing module"
clusto team assign <TEAM_ID> --prompt "Fix the flaky login test" --agent codex
clusto team wait <TEAM_ID> --timeout 600
clusto team merge <TEAM_ID> # preview
clusto team merge <TEAM_ID> --apply # merge worker branchesTeams persist across daemon restarts and are visible in the Teams view and in the Agents tab beside the chat or session that started them. The full command list is in the CLI reference, and the coordination model in Teams Autopilot.
Spread sessions over several machines
When one box runs out of CPU or memory, add another. Up to about five nodes form a leaderless cluster and the app lists sessions from all of them. A project can be pinned to a node, and clusto team create --node <NODE_ID> runs a team on another machine (node ids come from clusto cluster-status). Each session stays on the node that started it.
Tips that save time and tokens
- One task per worktree. Two agents in the same files will overwrite each other whatever tool you use.
- Keep tasks independent. Parallel sessions help with separate modules, reviews and investigations, less with a chain of dependent steps.
- Watch your plan limits. Every session signed in to the same account draws on the same usage. Clusto's chat status line shows the 5-hour and weekly limits.
- Review before merging. Preview with
clusto team mergebefore--apply, or check each branch in the git panel.
FAQ
How many Claude Code sessions can I run at once?
There is no fixed number for separate sessions; CPU, memory and your plan's usage limits set it. A Clusto team runs at most five active workers and queues the rest. Anthropic suggests 3 to 5 teammates for agent teams.
Can two Claude Code sessions work in the same repository?
Yes, if each has its own git worktree. Use git worktree add, claude --worktree <name>, or a Clusto team, which creates one per worker.
What is the difference between subagents, agent teams and parallel sessions?
Subagents report back to one main agent. Agent teams are several Claude Code instances coordinated by a lead inside one session. Parallel sessions are independent claude processes that you or a session manager run side by side.
Can I run Claude Code and Codex in parallel?
Yes. In Clusto, terminal sessions run any CLI, chats support Claude and Codex as providers, and a team can mix Claude and Codex workers.
Is Clusto free?
Clusto requires a license. The Personal plan is free for one user on up to three devices, and split panes are included. See Accounts and Licensing.