wmux Agent Workspace

Run CLI agents side by side on your machine, with worktrees, browser tools, and phone-based replies.

Stars
★ 418
Last updated
today
License
MIT
Primary language
TypeScript

At a glance

How it runs
Desktop appSelf-hosted serviceCLI
Works with
Universal · cross-platformCodex · Claude Code
Setup effort
Medium · a few setup steps
You'll need
Node 18+Python 3.xGitClaude Code, Codex CLI, Gemini CLI, or another agent CLI (installed and signed in separately)Shell / CLINetwork accessLocal filesystemMCP Server
Typical use
Developers exploring several implementations can fan one prompt into separate worktrees and review each change before adopting it.
Not a fit if
  • Users who need a native Intel Mac build
  • Teams relying on a stable Linux release
Source review
74/100 · Some gaps

What does this agent do, and when should you use it?

wmux is an Electron desktop workspace for Windows and Apple Silicon macOS, with split terminals for running CLI agents such as Claude Code, Codex, and Gemini. A separate background daemon owns terminal sessions, so they can survive app closure and be recovered after a restart. Users can fan a prompt out into up to eight tasks, each on its own Git worktree and branch, then review and adopt changes hunk by hunk. Agents can use MCP tools to operate a built-in browser or dedicated Chrome, and users can answer agent questions from a paired iPhone. The app runs on the user's machine; the iPhone client pairs directly with its daemon, keeping terminal output and replies off a cloud relay.

Install the wmux desktop app, open a workspace, and start installed CLI agents in real PTY terminals. A prompt can be split into up to eight tasks, each with a Git worktree and a fresh wtask/* branch; agents run in separate panes. Review completed diffs hunk by hunk and adopt selected changes with one git apply, or close the task or open a pull request. MCP tools let agents navigate, click, type, read snapshots, take screenshots, and use browser_repl and browser_replay. Agents can message other panes with send_message, while questions for the user appear in Fleet or the paired iPhone Inbox. The Git page shows issues, pull requests, checks, and failed logs, and supports actions such as commenting, approving, and merging.

  1. Developers exploring several implementations can fan one prompt into separate worktrees and review each change before adopting it.
  2. Engineers managing multiple repositories or workspaces can use Fleet to track agents that are running, waiting for replies, or ready for review.
  3. Teams asking agents to complete web workflows can use MCP tools with the built-in browser or dedicated Chrome.
  4. Developers who step away from their desk can pair the daemon with the iPhone app and reply to agent questions from its Inbox.
  5. Windows developers using Claude Code, Codex, or other CLIs can run them in native terminal panes without WSL.

How do you install or deploy this agent?

On Windows, install with winget or Chocolatey. On macOS, use an Apple Silicon Mac, download the DMG, and drag wmux into Applications. The iPhone app is available from the App Store; experimental Linux AppImage, deb, and rpm builds are on the releases page. For source builds, the README specifies Node 18+, Python 3.x, and the platform's native toolchain.

How do you use this agent?

Install on Windows:

winget install openwong2kim.wmux

Or use Chocolatey:

choco install wmux

Start the development app from source:

git clone https://github.com/openwong2kim/wmux.git
cd wmux
npm install
npm start

To pair an iPhone, open Remote → Share & pair in wmux, enable HTTPS over Tailscale, and scan the QR code with the iOS app. Open a workspace and start a CLI agent that you have installed and signed in to.

What are this agent's strengths and limitations?

Pros
  • Provides native terminal sessions on Windows and macOS; Windows users do not need WSL.
  • A background daemon keeps terminal sessions alive after app closure and offers recovery after crashes or reboots.
  • Supports up to eight parallel worktree tasks with hunk-level review and all-or-nothing adoption of selected changes.
  • Offers MCP controls for both built-in browser panes and a dedicated Chrome backend, with visible browser activity.
  • Pairs with an iPhone app for terminal output and replies directly between the phone and the local daemon.
Limitations
  • The current macOS build supports Apple Silicon only; Intel Mac builds are not produced.
  • Linux packages are explicitly experimental, so the README does not establish Linux as a stable release target.
  • Users must install and sign in to agent CLIs separately; wmux does not provide the models or subscriptions.
  • Building from source requires Node 18+, Python 3.x, and native build tools for Windows or macOS.
  • The README does not state the app or iOS client's price.

How does this agent compare with similar options?

The README names tmux as an inspiration and says wmux is not a tmux port: it is an Electron desktop app that adds agent workflows, Git worktrees, and browser features.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
wmux Agent Workspace This agent 74 · Some gaps Desktop app ★ 418 today TypeScript Codex · Claude Code
Superset 52 · Major gaps Desktop appFreemium ★ 15k today TypeScript Codex · Claude Code
Bernstein 78 · Good CLIFree + model costs ★ 1.2k today Python Codex · Claude Code · OpenAI API
Superagent 58 · Major gaps Desktop appFree + model costs ★ 29 today TypeScript Codex · Claude Code

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Some gaps
74/ 100 5-point scale 3.7 / 5
Trust 18/29
Reliability 9/14
Adaptability 14/18
Convention 16/18
Effectiveness 12/13
Verifiability 5/8
Why each dimension lost points
Trust18 / 29 · 3.1/5

The security documentation and README describe token-authenticated IPC, approval gates for cross-agent execution, a read-only web service bound to loopback by default, and sealed envelopes for phone push; the workflow also limits permissions, triggers, and token budget. These support the least-privilege, confirmation, and external-effects scores. Deductions: these are mainly documentation and workflow configuration, with insufficient implementation detail to assess the product’s full data flows, sensitive-data handling, or rollback behavior. Dependencies are listed with a few overrides, but there is no audit or vulnerability-remediation evidence.

Reliability9 / 14 · 3.2/5

The README describes daemon-owned persistent sessions, recovery, and all-or-nothing git apply for worktree changes. Install troubleshooting gives concrete errors and remedies, while CI files explain expected behavior and limitations. Deductions: the CI explicitly says the cross-platform baseline is not a release gate and that some platform branches have not been executed; repo-wide lint is advisory. Static evidence cannot establish runtime robustness of error handling or dependency availability.

Adaptability14 / 18 · 3.9/5

The README covers multiple CLI agents, workflows, and Windows/macOS users, and describes browser, phone, worktree, Git, scheduling, and MCP profile scenarios, with some platform and feature boundaries. Deductions: the broad compatibility claims are not backed by equally specific adaptation details for every environment and agent; trigger and applicability boundaries are described across feature sections rather than in a consistent, comprehensive specification.

Convention16 / 18 · 4.4/5

The documentation is clearly organized, with entry points for installation, building, troubleshooting, FAQ, limitations, and detailed guides; the MIT license text is explicit. The README gives platform and installation guidance, while package.json provides version, engine, and dependency information. Deductions: the supplied material has no changelog or durable maintenance commitment; release paths are described, but responsibility and maintenance commitments are not clear. There is no formal stability policy for names or components.

Effectiveness12 / 13 · 4.6/5

The README describes concrete outputs for agents, including browser control, terminal use, agent-to-agent communication, task fan-out, and review workflows; worktree isolation and selective adoption provide clear potential user value. Deductions: cost, resource use, and realized benefit are not quantified, and the described experience is presented by repository materials rather than independently established evidence.

Verifiability5 / 8 · 3.1/5

Most feature, installation, and security claims point to repository documentation; the workflows document trigger constraints, permissions, and CI failure context. The README sometimes distinguishes defaults, experimental builds, and known limitations. Deductions: many product capabilities are asserted in the README, while the linked implementation documents and corresponding source are not included here for cross-checking; static evidence cannot verify the depicted or described runtime behavior.

Risks and how to mitigate them
  • This assessment uses only the supplied static material and cannot confirm runtime behavior of demos or security controls.
  • The README labels Linux builds experimental; the Windows installer uses a SignPath test certificate and may trigger an unknown-publisher SmartScreen warning.
  • The CI material says the cross-platform baseline does not gate releases and that some platform branches have not been exercised in CI.
Evidence confidence: Low Reviewed Oct 09, 2026 Reviewed revision cbab39888b4e
See the full review method →

FAQ

Does wmux include Claude, Codex, or Gemini usage credits?
No. wmux runs CLI agents that users install and sign in to themselves; users are responsible for following their AI provider's terms.
Do agents stop when I close the desktop app?
No. A standalone daemon owns terminal sessions. The README also describes recovery after crashes or reboots and configurable pane restart policies.
Does remote iPhone access relay terminal content through the cloud?
The README says terminal output and replies travel directly between the phone and the local daemon. Push notifications pass through the project's relay as sealed envelopes it cannot read.
Is Linux suitable for a stable daily deployment?
The project offers experimental AppImage, deb, and rpm builds; the README does not describe Linux as a stable supported release platform.
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents