Current Chrome Browser Skill for Claude Code

A safety-focused Claude Code skill for operating the Chrome window you already use through Microsoft's extension-backed Playwright MCP server.

It is designed for authenticated browser workflows where launching a clean automation profile is the wrong behavior. The skill requires Claude Code to select an existing tab, reason from accessibility snapshots, verify external side effects, and stop on login challenges, limits, or ambiguous submissions.

Features

• Uses only the current Chrome session and selected existing tab.
• Prefers accessibility snapshots and stable element references.
• Verifies account, destination, payload, preview, and result.
• Prevents silent fallback to another browser or profile.
• Stops on authentication challenges, platform limits, and ambiguity.
• Supports long autosaved editors with reload verification.

This repository contains an instruction layer, not browser automation software. It also includes an optional local companion CLI that records tab leases and guards external side effects. It requires the official Playwright MCP server and browser extension.

Prerequisites

• Claude Code CLI
• Node.js 20 or newer
• Google Chrome or Microsoft Edge
• Official Playwright MCP browser extension
• Official @playwright/mcp server configured with --extension

Configure Playwright MCP

Install the extension using the official Playwright documentation, then add the server to Claude Code.

Give the server a distinct name such as playwright-chrome. Do not call it playwright: plugins and marketplaces ship their own Playwright MCP servers, those run a clean launched profile, and their tools carry the same names as these. Two servers named alike are indistinguishable at the point of use, and picking the wrong one does not raise an error — it returns a blank page or a login screen from a profile that is not yours.

Windows:

claude mcp add playwright-chrome -- cmd /c npx -y @playwright/mcp@latest --extension

macOS or Linux:

claude mcp add playwright-chrome -- npx -y @playwright/mcp@latest --extension

Restart Claude Code and verify the connection:

claude mcp list

Use /mcp inside Claude Code to approve the server when prompted.

If another browser MCP server is already configured, remove or disable it for the duration of authenticated browser work. The skill can be told which server to trust, but it cannot recover the time lost to a wrong call that looked like an empty page.

Install The Skill

Personal installation makes the skill available in every project.

PowerShell:

$repo = Join-Path $env:TEMP "current-chrome-browser-skill"
git clone https://github.com/mikhailpetrusheuski/current-chrome-browser-skill $repo
$target = Join-Path $HOME ".claude\skills\current-chrome-browser"
New-Item $target -ItemType Directory -Force | Out-Null
Copy-Item (Join-Path $repo "SKILL.md") (Join-Path $target "SKILL.md") -Force

Bash:

git clone https://github.com/mikhailpetrusheuski/current-chrome-browser-skill /tmp/current-chrome-browser-skill
mkdir -p ~/.claude/skills/current-chrome-browser
cp /tmp/current-chrome-browser-skill/SKILL.md ~/.claude/skills/current-chrome-browser/SKILL.md

For one project, copy SKILL.md to:

<project>/.claude/skills/current-chrome-browser/SKILL.md

A project installation is visible only inside that project. In a multi-repository workspace, place it at the workspace root that Claude Code opens, not in a nested repository, or the skill will not be listed.

Restart Claude Code if .claude/skills did not exist when the current session started. Until the restart the skill is absent from the session, and asking for it reports an unknown skill rather than a pending installation.

Usage

Invoke the skill explicitly:

/current-chrome-browser list the tabs in my current Chrome and summarize the active documentation page

/current-chrome-browser use the current GitHub tab to prepare an issue, but do not submit it

/current-chrome-browser update the draft in the selected editor, wait for autosave, reload, and verify persistence

The skill has disable-model-invocation: true, so Claude cannot invoke it automatically. Access to an authenticated browser remains under explicit user control.

This is deliberate, and it has a confusing symptom: if Claude tries to invoke the skill on its own it is told the skill is unknown, which reads as a broken installation. Only the user typing /current-chrome-browser starts it. Ask the user to run the command rather than diagnosing the installation.

See EXAMPLES.md for more prompts and expected behavior.

Optional Companion Guard

The TypeScript CLI adds machine-checkable workflow state without controlling the browser or reading its session:

npm install
npm run build
node dist/cli.js preflight

It can bind work to a selected tab, require an explicit authorization flag, detect a repeated committed payload, and append confirmation records. Raw payload content is never written to its ledger. State stays in the ignored .browser-guard/ directory.

See GUARD.md for the command contract and integration workflow.

For connection failures, stale refs, large snapshots, shadow DOM, file uploads, parallel-agent interference, editor corruption, and ambiguous submissions, see TROUBLESHOOTING.md.

Security

The skill is not a sandbox. Playwright MCP can interact with authenticated websites available in the selected browser tab. Keep Claude Code confirmation prompts enabled and read SECURITY.md before using sensitive accounts.

Learn more in the official Claude Code skills documentation and Playwright MCP guide.

License

MIT