Nobrainer Browser

nobrainer-browser

Single-purpose skill: get four browser-automation tools installed and wired so that BOTH Claude Code and Codex CLI can drive a browser. Same conventions as nobrainer-paperclip-setup: stdlib-only, cross-platform paths, atomic writes, no personal data, designed to be published publicly.

Triggers

Activate when the user says:

• nobrainer-browser, "install browser tools", "set up browser automation"
• "install playwright mcp", "wire playwright mcp", "playwright mcp for claude"
• "install webwright", "install agent-browser"
• PL: "zainstaluj playwright mcp", "skonfiguruj browsery dla AI", "zainstaluj narzedzia do automatyzacji przegladarki"

What it does

1. Detects what is already installed (Node, npm, pip, each of the four tools, and the wiring state of Claude/Codex config files).
2. Installs each tool through its native channel:
   • Webwright: two modes.
     • marketplace (default, recommended): no clone, no pip, no playwright download. The host's plugin loader fetches microsoft/Webwright from GitHub. Supports claude and codex hosts only.
     • source (opt-in via --mode source): git clone into a skill-managed directory, pip install --user -e <clone>, then python -m playwright install chromium. Required for openclaw and hermes hosts. Hosts: claude, codex, openclaw, hermes (select via --hosts a,b,c; default claude,codex). For Claude the skill prints the slash commands the user must run; for Codex/OpenClaw it shells out; for Hermes it creates a symlink under ~/.hermes/skills/webwright.
   • Playwright MCP: no global install; resolves a 7-day-safe version via npm view @playwright/mcp time --json, caches the pin under state.json, and exposes it for wiring.
   • Playwright CLI: npm install -g playwright@<safe-version> followed by playwright install (Chromium + Firefox + WebKit). On Linux, --with-deps is supported via a flag.
   • agent-browser: npm install -g agent-browser@<safe-version> by default (also supports --method=cargo and --method=brew), then agent-browser install for Chrome for Testing, then npx skills add vercel-labs/agent-browser for the discovery stub.
3. Wires the Playwright MCP server into both editors:
   • Claude: tries claude mcp add playwright npx "@playwright/mcp@<pin>"; falls back to an atomic edit of ~/.claude.json (mcpServers.playwright).
   • Codex: tries codex mcp add playwright npx "@playwright/mcp@<pin>"; falls back to a conservative splice of [mcp_servers.playwright] into ~/.codex/config.toml that preserves every other section.
4. Verifies. Each tool exposes a --version probe; the MCP wiring is verified by reading the config files back.

Architecture

nobrainer-browser/
|- SKILL.md
|- README.md
|- pyproject.toml
`- nobrainer_browser/
   |- __init__.py
   |- __main__.py            # argparse + dispatch (detect/install/wire/unwire/verify/status)
   |- paths.py               # cross-platform user/data/log dirs + atomic writes
   |- npm_safe.py            # 7-day npm registry cooldown (resolve + install + npx)
   |- detect.py              # host + per-tool + wiring snapshot
   |- verify.py              # health checks
   |- tools/
   |  |- webwright.py        # git clone + pip install -e + playwright install chromium
   |  |- playwright_mcp.py   # resolve + cache safe version (no global install)
   |  |- playwright_cli.py   # npm_safe.install_global + playwright install
   |  `- agent_browser.py    # npm/cargo/brew + agent-browser install + skills add
   `- wire/
      |- claude.py           # `claude mcp add` + ~/.claude.json fallback
      `- codex.py            # `codex mcp add` + ~/.codex/config.toml fallback

CLI surface

Tool-to-editor wiring

Webwright supported hosts

Webwright can be registered with four host environments. The marketplace mode is the default and works without any local checkout; the source mode is required for hosts that only accept a local path.

Security model

• Supply-chain cooldown. Every npm install resolves to the latest version published >= 7 days ago. NPM_SAFE_BYPASS=1 disables the cooldown for the current process and emits a loud stderr warning. The cutoff defaults to 7 days; override with --cooldown-days.
• Atomic writes. All config-file edits go via tempfile.mkstemp in the destination directory + replace, with Windows-aware retry on PermissionError. No partial writes.
• No secret logging. API keys are detected (truthy/falsy) but never printed, persisted, or written to logs.
• No personal data. The package contains no usernames, emails, hostnames, or company-specific paths. Safe to publish publicly.
• Shell-free subprocess. Every external invocation uses shell=False and resolves the executable via shutil.which before running.

For LLMs

Invoke this skill when:

• The user asks to set up browser automation for an AI agent.
• The user mentions Playwright MCP, Webwright, or agent-browser by name.
• Either Claude Code or Codex CLI needs to gain browser control and the current config files don't already list a Playwright MCP server.

Do NOT invoke for:

• Running an existing Playwright test suite (that's the test runner, not this skill).
• Day-to-day browsing actions performed through an already-wired MCP server.
• Cleaning up test data inside a specific web app (use the appropriate domain skill instead).

Requirements

• Python 3.10+
• Node.js >= 20 (for npm/npx)
• git (for Webwright clone)
• macOS, Windows 10+, or any modern Linux

Optional:

• Rust toolchain (cargo) if using --method=cargo for agent-browser.
• Homebrew (brew) if using --method=brew for agent-browser.
• One of OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY for Webwright runtime.

Source: nobrainer-tech/nobrainer-claude-skills — distributed by TomeVault.