A control room for terminal coding agents
Claude Squad addresses a problem that appears as soon as coding agents become useful: one session is easy to supervise, while five terminals are not. The application lists agent sessions in one terminal interface, shows captured output, and provides a diff view for the selected task. A developer can attach to reprompt an agent, pause or resume work, review changes, commit, and push without hunting through terminal tabs.
Each session runs under tmux in its own Git worktree and branch. That is the right isolation boundary for parallel source changes. Agents do not switch the shared checkout under one another, and their uncommitted files stay separate. Profiles allow the same dashboard to launch Claude Code, Codex, Gemini, Aider, or another shell command, making Claude Squad a supervisor rather than a Claude-only wrapper despite its name.
The tool is most effective when tasks are independent: one agent can fix a test, another update documentation, and a third investigate a separate bug. Closely coupled edits still create merge conflicts and require human coordination.
The interface covers the essential loop
Creating a session with n opens an agent, while N supplies an initial prompt. The list and preview provide quick status. Attaching hands the terminal to the selected session, and a detach chord returns to the dashboard. The diff tab makes review part of the normal navigation rather than a final surprise. Checkout commits and pauses a session; another action commits and pushes the branch through GitHub.
Named profiles are a practical addition. A configuration can expose separate Claude, Codex, and local Aider commands, then present them in a picker for each new task. This is better than changing one global command whenever a different model or provider fits the job. It also means profile strings are executable shell commands, so shared configuration needs code review and careful quoting.
Experimental auto-yes accepts prompts for supported agents. That can keep background tasks moving, but it removes a crucial intervention point. Agent approval screens often guard shell execution, file edits, or network access. Automatically agreeing makes the agent's sandbox and OS permissions the only remaining limit. Use it only in disposable environments with narrow credentials and a clean recovery path.
The worktree is not the whole environment
Installation through Homebrew or the shell script is short. Claude Squad additionally requires tmux and the GitHub CLI, plus a working agent command and provider credentials. It must start inside a Git repository, and current session creation is tied to that repository. An open feature request asks to create a session in another project without quitting and relaunching.
A new Git worktree contains tracked files, not the complete development environment. Ignored .env files, package directories, virtual environments, compiled caches, and local configuration are missing. Issue 260 also notes that parallel web servers can all choose port 3000 and Docker Compose sessions can share a project name. Proposed setup-hook pull requests address these problems, but they are not documented features of v1.0.19. Today, teams need their own bootstrap script or tools that derive per-worktree ports, copy safe configuration, and install dependencies.
This overhead can erase the speed gained by parallelism in a large monorepo. It can also create dangerous shortcuts, such as copying production secrets into every agent directory. Prefer generated local credentials and explicit test services.
State and cleanup need immediate attention
The most important current report concerns state.json. Pull request 317 says each TUI and the auto-yes daemon loads one global instance list, then later writes its complete in-memory copy without locking or merging. Two Claude Squad processes can therefore overwrite one another's sessions. Until a released fix adds coordination, run a single process and back up any important branches outside the application's state.
Cleanup has a sharper edge. Pull request 307 says cs reset permanently deletes all sessions and worktrees without asking, including uncommitted changes. The patch adds a warning and force flag, but users of the current release should treat reset as destructive and inspect every worktree first. Branch commits and remote pushes provide recovery that the dashboard cannot.
Windows support is not ready in the published native form. Issue 275 reproduces the Windows binary launching its interface but failing on every new session because the chosen PTY library returns unsupported. Another report shows problems with alternative tmux implementations. WSL may provide the expected Unix stack, but that is a different environment and should be tested explicitly.
Other workflow races remain visible. Issue 266 says an initial prompt can be sent before Codex or another CLI is ready, leaving the session waiting with no prompt. A proposed fix detects program-specific prompt markers, which itself shows how supporting arbitrary agent commands becomes fragile as their terminal UIs change.
Health and the buying decision
Version 1.0.19 shipped on June 17, 2026, and the repository was pushed on July 30. The 53 open issues and pull requests include significant community work on recovery, state locking, focus mode, editor launch, worktree hooks, Windows, and multi-repository use. Activity is real, but many useful fixes are still proposals rather than released behavior.
The README is concise and covers installation, profiles, keybindings, prerequisites, and the tmux-plus-worktree architecture. It needs stronger warnings around reset, concurrent state, ignored files, and native Windows.
Claude Squad is worth trying when manual terminal sprawl has become the bottleneck. Its dashboard and isolation model are immediately understandable. Keep the human in the loop, run one instance, create durable branches early, and regard every convenience around automatic approval, pushing, and reset as privileged until the current safety work lands.