PlugWiki

cmux-browser Skill (cmux) for Claude Code: What It Does and How to Install It

Makes the agent drive the browser inside cmux, a macOS terminal for coding agents, without stealing focus.

Installs
8.9K
on skills.sh
Stars
28.6K
on GitHub
Files
12
besides SKILL.md
Updated
Nov 2026
last repo commit

The agent uses the cmux browser command. It opens a browser surface, takes an interactive snapshot that gives each element a reference such as e1, then fills, clicks and waits for text, a selector, a URL part or the load state. Every command on an existing page carries an explicit surface handle, and SKILL.md says never to guess it from what has focus.

It needs the cmux app, which the README says is macOS only, and jq for the discovery snippets. cmux uses WKWebView, so trace and screencast recording, network interception and raw input injection return not_supported. Saved login state holds cookies and storage, so treat it as sensitive.

Good for

  • cmux users who run Claude Code or Codex and want the agent to use the built-in browser
  • Developers checking a page, filling a form or extracting data from a site
  • Anyone writing shell scripts around cmux browser commands

Install cmux-browser

Claude Code

  1. 1

    Copy this prompt.

    Install the cmux-browser skill in Claude Code. Run this command:
    
    npx --yes skills@1.5.23 add manaflow-ai/cmux --global --yes --skill cmux-browser --agent claude-code codex --copy
    
    Then check that it loaded and tell me how to use it.
  2. 2

    Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

Install the cmux app first (the README documents a DMG, which auto-updates, and a Homebrew cask), because the skill only drives the browser inside cmux, which is macOS only. The skill install command is the one SKILL.md gives under Skill distribution and refresh. It needs jq for the surface-discovery snippets and runs local commands, so there is no Claude app upload route.

From the project README. Source.

What's in the skill

When it loads

End-user browser automation with cmux. Use when you need to open sites, inspect or interact with browser surfaces, wait for page state, and extract data without stealing focus.

Sections in SKILL.md

  1. Read the CLI contract first
  2. Find an existing browser surface without changing focus
  3. Core workflow
  4. Wait
  5. Viewport sizing (WKWebView)
  6. Limits (WKWebView)
  7. Troubleshooting js_error
  8. Skill distribution and refresh
  9. Deep-dive references
  10. Ready-to-use templates

Key facts

Installs on skills.sh
8,885 (Nov 2, 2026)
Last repo commit
Nov 1, 2026

What it includes

  • cmux browser --help and cmux --version before any exact command

    Runs code · Read the CLI contract

  • cmux identify --json and cmux tree --all --json list browser surfaces

    Both are read-only listings per SKILL.md, and neither changes focus. The output can hold page URLs and titles.

    Runs code · Find a surface

  • cmux --json browser open <url> --focus false returns a surface ref

    Runs code · Open a browser surface

  • snapshot --interactive, then fill e1 "text" and click e2 --snapshot-after

    Refs go stale, so snapshot again after navigation or a modal.

    Try: “Open example.com in cmux and fill in the contact form”

    Runs code · Snapshot, fill and click

  • wait --selector, --text, --url-contains, --load-state and --function with a timeout

    Runs code · Wait

  • viewport <w> <h>, viewport reset, and download list for saved files

    Runs code · Viewport, zoom and downloads

  • form-automation.sh, authenticated-session.sh and capture-workflow.sh

    capture-workflow.sh writes a snapshot text file and a screenshot file. authenticated-session.sh loads and saves a state file.

    Runs code · Templates

  • Fall back to get text body or get html body when snapshot fails

    Runs code · Troubleshoot js_error

From the official docs

Permissions

Medium risk

The agent can click and submit in a real browser, including logged-in pages, and can save session state to disk. Use throwaway accounts first and keep any state file private.

What it can do

  • Read your data
  • Create and change things
  • Delete things
  • Read-only mode

Watch out for

  • It can act in logged-in pages

    The authentication guide covers login, OAuth and 2FA flows, cookie setting and state save and load. Only let it use accounts you are happy for an agent to act in. github.com

  • Saved state and cookies hold your session

    references/authentication.md says saved state contains cookies, localStorage, sessionStorage and open tab metadata, and that state files hold auth tokens and must never be committed. Its examples keep them in a mode 700 directory with chmod 600 files, but the authenticated-session.sh template defaults to ./auth-state.json in the working directory. The guide also shows cookies set with a value from an environment variable, and cookies clear --all plus deleting the state file after sensitive tasks. AGENTS.md calls cookies and saved state from authenticated pages sensitive. github.com

  • Listings can expose page URLs and titles

    SKILL.md says raw tree and list output can contain page URLs and titles, so filter or redact it before logging or pasting. github.com

  • The install command refreshes global skills

    The pinned skills 1.5.23 command uses --global --yes --copy for claude-code and codex, which writes the full skill into both global skill folders. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details
Turning on read-onlyNo read-only switch exists.
Sign-in and keysSaved state files contain cookies, localStorage and sessionStorage; the guide says to keep them in a private directory.
Scopes
  • Controls the cmux browser: opens pages, clicks, fills forms and reads page content

  • Can save and load browser state with cookies, localStorage and sessionStorage

  • Runs shell templates that write snapshot, screenshot and state files

Questions

What does the cmux-browser skill do?

It teaches the agent to use the cmux browser command. The agent can open a site, read the page structure, click and fill elements, wait for page state and extract data, all inside cmux.

How do I install it?

Use the install box at the top of this page. Install the cmux app first, then add the skill with the command SKILL.md gives, which targets both Claude Code and Codex. The box lists the folder routes too.

Does it work without cmux?

No. The skill only drives the browser inside cmux, and the cmux README says the app is macOS only for now.

Is it official?

Yes. It lives in the manaflow-ai/cmux repository, the vendor's own repo.

Does it need an API key?

No key is named in SKILL.md. Login examples in the reference guides pass a password from an environment variable such as APP_PASSWORD.

Is it safe?

It can click, fill and submit in a real browser, and it can save and load browser state that holds cookies and storage. AGENTS.md says to treat URLs, titles, snapshots, cookies and saved state from authenticated pages as sensitive. Keep state files out of git.

What can it not do?

SKILL.md says trace and screencast recording, network route interception and mocking, and low-level raw input injection return not_supported in cmux's WKWebView browser.

Information

updated Nov 2, 2026

Folder
skills/cmux-browser

manaflow-ai/cmuxskills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 6
  1. cmux-browser SKILL.md · github.com
  2. cmux README · github.com
  3. cmux-browser AGENTS.md · github.com
  4. cmux-browser references/authentication.md · github.com
  5. capture-workflow.sh · github.com
  6. authenticated-session.sh · github.com

Updated Nov 2, 2026.