mrkeyoor.com_
Sat 26 Sept 20:01 UTC
Automationevaluationupdated 26 Aug 2026

claude-squad review

Claude Squad is a terminal interface for running several coding-agent sessions at once. It puts each Claude Code, Codex, Gemini CLI, Aider, or custom command in its own tmux session and Git worktree, then lets you inspect diffs and move finished work back into your branch.

+27stars / 7d
Verdict

Our Claude Squad run installed 48 Go packages, built successfully, and passed all 6 tests in 91 seconds combined. It is worth using on macOS or Linux when you already know how to manage worktrees and want one screen for several agent CLIs. Skip it for native Windows, automatic environment provisioning, or hands-off prompt dispatch until the documented gaps are fixed.

We ran it

Lab card: what happened when we ran claude-squadScreenshot of claude-squad (smtg-ai.github.io/claude-squad)
Install✓ · 21s48 packages
Build✓ · 26s
Tests✓ · 44s6 passed · 0 failed of 6 (go test)
Repo86 files~10,088 lines of source · 1.5 MB · 5 CI workflows

Answers from our run

Does claude-squad build from source?

Dependencies installed in 21 seconds (48 packages), and the build succeeded in 26 seconds. We cloned commit ce1ffb4 into a clean Debian container with 3 CPUs and no project-specific setup.

Do claude-squad's tests pass?

Yes: 6 of 6 passed when we ran the project's own test command (go test). Some failures need services or credentials a bare container does not have.

Who should not use claude-squad?

Native Windows users: open issue 275 says the published v1.0.17 Windows binary cannot create a session because its PTY dependency is unsupported there.

What are the alternatives to claude-squad?

Claude Code, Codex CLI, tmux. Our Claude Squad run installed 48 Go packages, built successfully, and passed all 6 tests in 91 seconds combined.

Setup4/5Source build passed; tmux, gh, agents, and worktree setup remain
Docs4/5Clear install, keys, profiles, controls, and worktree explanation
Community4/58,367 stars with August 2026 releases and issue activity
Maturity4/5v1.0.20 fixes recovery, but Windows and startup races remain

Discussed on

  1. hnClaude Squad – A terminal app that manages multiple Claude Code instances5 points

Who it’s for

Developers who can supervise several independent coding tasks from one terminal.
Claude Code and Codex users comfortable with Git worktrees, tmux, branches, and reviewing agent output.
Teams that want local isolation between agent sessions without adopting a hosted orchestration service.

Who it’s NOT for

Native Windows users: open issue 275 says the published v1.0.17 Windows binary cannot create a session because its PTY dependency is unsupported there.
Repositories that require copied secrets, unique ports, separate databases, or per-worktree dependency setup with no custom scripting: open issue 260 requests the missing setup hook for exactly these cases.
Anyone expecting unattended prompts to be reliable across every supported CLI: issue 266 reports that Codex can lose a startup prompt because Claude Squad sends it before the CLI is ready.
Teams unwilling to review agent actions: the README labels auto-accept as experimental, and that mode approves prompts for Claude Code and Aider.

Setup reality

Our source install succeeded in 21 seconds and fetched 48 Go packages. The build passed in 26 seconds. Tests finished in 44 seconds with 6 passed and 0 failed of 6. The 1.5 MB checkout contained 86 files and about 10,088 source lines.

Daily use also needs tmux, GitHub's gh CLI, Git, and at least one installed agent command. Hosted agents need their usual login or API key; the README specifically tells Codex users to set OPENAI_API_KEY. Claude Squad stores profiles in a local JSON config.

Each session creates a Git worktree and branch. Untracked environment files and installed dependencies do not automatically follow it, and parallel servers can fight over ports. Our scan found 5 CI workflows but no Dockerfile or tests directory.

Six passing tests support the small Go core

Claude Squad gives parallel coding agents a shared terminal dashboard. Each task runs under tmux in a separate Git worktree, so two agents can edit the same repository without writing into the same checkout. The interface shows sessions, terminal output, and diffs. From there you can attach to an agent, pause it, resume it, check out its work, or commit and push its branch. The design is understandable because it builds on tools developers already use.

Our checkout was compact: 86 files, about 10,088 source lines, and 1.5 MB. The install fetched 48 Go packages in 21 seconds. Building took 26 seconds, then the test step passed 6 of 6 tests in 44 seconds. That result makes the source itself easy to trial. It says less about whether five agents can run your application's databases, ports, and dependency trees at the same time.

Worktrees prevent file conflicts but do not clone your environment

A Git worktree isolates tracked files and branch state. It does not include ignored .env files, a Python virtual environment, node_modules, local databases, or a unique port allocation. Open issue 260 describes new Claude Squad sessions starting without those pieces and asks for a setup hook. The reporter also calls out Docker Compose project-name collisions and multiple servers trying port 3000.

Plan for that gap before creating several sessions. A repository-specific script can copy safe configuration, install or link dependencies, assign ports, and create disposable databases. Keep real secrets out of the worktree when an agent does not need them. Claude Squad's isolation is useful, but it only covers Git. The 48-package Go install does not remove application-level shared state.

What happened when we ran it

Our sandbox installed Claude Squad in 21 seconds, adding 48 Go packages. The build completed in 26 seconds. Tests took 44 seconds and reported 6 passed, 0 failed of 6. We ran commit ce1ffb4 in an unprivileged Debian container with 3 CPUs, 8 GB of RAM, and no secrets.

The repository had 5 CI workflow files, no Dockerfile, and no tests directory. Go tests can live beside package code, so the missing directory does not conflict with the 6 passing tests. We did not launch Claude Code, Codex, Gemini, or Aider; authenticate to GitHub; create a real worktree; push a branch; or measure concurrent agent resource use. Those operations need user accounts and a repository with meaningful tasks.

tmux, gh, and an agent login are still prerequisites

The README requires tmux and GitHub's gh command. The default launched program is claude, and maintainers recommend its latest version. Codex users are told to export OPENAI_API_KEY, while other programs keep their own authentication rules. Named profiles in ~/.claude-squad/config.json can select commands such as claude, codex, or an Aider invocation with a local model.

That command flexibility is useful and risky. A profile is a shell command Claude Squad starts inside the worktree, so configuration deserves code-review discipline. The experimental --autoyes option automatically accepts prompts for Claude Code and Aider. Use it only in a disposable branch with narrow credentials and a clear task. A separate worktree can contain file damage, but it cannot undo an agent's network calls, cloud changes, or messages sent with available credentials.

Codex startup prompts can race the CLI

Open issue 266 reports that the prompt supplied through the new-session flow may be written before Codex has finished initializing. The described result is a blank waiting prompt: the intended task never arrives. The report traces the path to a direct PTY write and notes that readiness detection exists for some assistants but not Codex. That is a real concern for anyone launching tasks and immediately moving on.

Until the project confirms a fix, watch each new session long enough to see that the agent received its instruction. Release v1.0.20, published August 20, 2026, fixed recovery when a tmux session dies and corrected base-commit recording for sessions created from existing branches. Those changes show maintainers working on session state, but the release notes do not claim to solve issue 266.

Native Windows is a published binary without a working session path

Issue 275 reports that the v1.0.17 Windows binary launches its interface but fails as soon as the user creates a session. According to the report, the creack/pty dependency compiles through an unsupported stub on Windows and returns an error before tmux is invoked. The README does not state native Windows support limits. WSL may change the environment, but the issue says there is no native workaround for the tested binary.

GitHub showed 8,367 stars, 53 combined issues and pull requests, and a last push on August 20, 2026. v1.0.20 was released the same day, while issue 260 had activity on August 14. That is healthy current motion, paired with visible platform and orchestration gaps. On macOS or Linux, Claude Squad is a good supervisory layer for developers who review every branch. It does not yet make parallel agents automatic or consequence-free.

Alternatives

ProjectWhat it isPick it when
Claude Code gh↗Anthropic's coding agent includes its own subagents and terminal workflow without a separate session dashboard.pick this instead when one Claude session with delegated subagents is enough and separate Git worktrees would add friction.
Codex CLI gh↗OpenAI's local coding agent can work directly in a repository and run parallel agents through its own workflow.pick this instead when Codex is your only agent and you do not need one TUI spanning several providers.
tmux gh↗A terminal multiplexer that can host several agent commands without imposing branch or checkout behavior.pick this instead when you want complete control over sessions and are willing to manage worktrees and diffs yourself.

What people are saying

  1. [github-trending] smtg-ai/claude-squad

Sources

  1. Claude Squad README
  2. Claude Squad v1.0.20 release
  3. Worktree environment setup hook request
  4. Native Windows PTY failure report
  5. Codex startup prompt race report

More automation reviews

runner-images · agent-fleet-manager · kargo · Rose · alchemy · laya · the whole board →