Playwright CLI vs MCP vs Claude in Chrome: Which Browser Automation Tool Should You Use?

In 2025, browser automation stopped being a developer-only concern. AI coding agents — Claude Code, GitHub Copilot, Cursor, Windsurf — needed to see and interact with web pages as part of their workflows. Opening a localhost preview, filling out a form, verifying a visual change, debugging a console error — these became agent tasks, not human tasks.

The problem: traditional automation tools like Playwright and Selenium were designed for deterministic test scripts, not for AI agents making decisions in real time. The accessibility tree of a complex page can consume 50,000+ tokens in a single snapshot — burning through context windows and API budgets.

Three tools emerged to solve this from different angles. Microsoft released Playwright MCP in March 2025 — an MCP server exposing Playwright's full browser control to any AI agent. Then in February 2026, they shipped Playwright CLI (@playwright/cli) — a leaner, token-efficient alternative built for coding agents. Meanwhile, Anthropic launched Claude in Chrome — a Chrome extension that lets Claude control your actual browser, with all your logged-in sessions intact.

Playwright MCP was the first major bridge between AI agents and browser automation. It implements the Model Context Protocol (MCP) — the open standard for connecting AI models to external tools — and exposes 25+ browser control tools that any compatible client can call.

The core innovation is accessibility snapshots instead of screenshots. When an agent calls browser_snapshot, Playwright captures the page's accessibility tree — the same structured representation that screen readers use — and returns it as text. Each interactive element gets a unique ref identifier. The agent reads the text, decides what to do, and issues commands like browser_click with a ref.

This is faster and more deterministic than vision-based approaches. No image encoding, no pixel coordinate guessing. But there is a cost: every snapshot dumps the entire accessibility tree into the conversation context. On a content-rich page, that is thousands of tokens per interaction. Across a 12-step workflow, conversations can balloon past 90,000 tokens of stale page data.

Key strengths include maturity (1+ year in production), cross-browser support (Chromium, Firefox, WebKit, Edge), Docker support for CI/CD, device emulation, and proxy routing. Limitations include high token consumption (~114K per typical task), 25+ tool schemas loaded into context even when unused, noisy accessibility trees on complex pages, and manual authentication setup per session.

Playwright CLI (@playwright/cli) launched in February 2026 as Microsoft's answer to the token efficiency problem. The insight: CLI invocations are inherently cheaper than MCP tool calls because they do not require loading large tool schemas or streaming full accessibility trees into the model's context window.

Instead of 25+ MCP tools, the CLI exposes a single binary with subcommands: playwright-cli open, playwright-cli click <ref>, playwright-cli snapshot. Snapshots are saved to disk as YAML files rather than injected into conversation context. The agent reads only what it needs, when it needs it.

The result: Microsoft's own benchmarks show ~27,000 tokens per typical task vs ~114,000 for MCP — a 4x reduction. For teams paying per-token for AI API calls, this is a significant cost difference at scale.

The CLI also introduces named sessions — you can run multiple browser instances simultaneously with playwright-cli -s=session1 open and switch between them. A built-in monitoring dashboard (playwright-cli show) provides live screencasts of all running sessions. Additional features include full cookie, localStorage, and sessionStorage CRUD commands, network mocking via route commands, and video/trace recording.

Limitations: it is very early (v0.1.1 with alpha Playwright dependency), has no test execution capability (browser control only), is headless by default, and sessions are in-memory unless you use the --persistent flag.

Claude in Chrome takes a fundamentally different approach. Instead of launching a separate browser instance, it controls your actual Chrome browser — with all your cookies, login sessions, extensions, and browser fingerprint intact.

The architecture uses Chrome's native messaging protocol to bridge three layers: the Chrome extension (running in-browser), a native messaging host (on disk), and Claude Code or Claude Desktop (as the MCP client). When connected, Claude can navigate, click, type, read page text, execute JavaScript, capture screenshots, and record GIFs — all in your visible browser window.

The killer feature is authentication inheritance. Claude can interact with your Gmail, Google Docs, Notion, Jira, CRM — any app you are logged into — without any API keys, OAuth flows, or credential injection. This makes it uniquely powerful for tasks like "check my calendar, then draft a response to that Slack thread based on the meeting notes in Google Docs."

For developers, the most common workflow is build-then-verify: Claude writes code, opens the localhost preview in Chrome, takes a screenshot, checks for visual regressions or console errors, and iterates — all without the developer leaving their editor.

Key strengths include using your real browser sessions (no re-authentication), low bot detection risk, targeted responses that do not dump full page trees, GIF recording for documentation, and native pixel-level computer use control. Limitations include Chrome and Edge only (no Firefox/WebKit), no headless mode, unsuitability for CI/CD, JS alert/dialog blocking, paid Anthropic plan requirement, and an 11.2% prompt injection risk even with mitigations.

Token efficiency is the biggest differentiator. Playwright CLI uses ~27,000 tokens per task by saving state to disk as YAML. Playwright MCP uses ~114,000 tokens by streaming full accessibility trees into context. Claude in Chrome falls somewhere in between, using targeted responses rather than full page dumps.

Architecture: CLI works via shell commands with zero tool schema overhead. MCP loads 25+ tool schemas into the agent's context. Chrome loads ~18 tools and connects via native messaging to a real browser.

Browser support: Both Playwright tools support Chromium, Firefox, WebKit, and Edge. Claude in Chrome is limited to Chrome and Edge.

Authentication: CLI uses persistent profiles or manual login. MCP uses storage state files. Chrome inherits your real browser sessions — the only tool that avoids the authentication problem entirely.

CI/CD readiness: CLI and MCP are both CI-ready (MCP has Docker support). Chrome requires a visible browser and is not designed for automated pipelines.

Element targeting: CLI and MCP both use snapshot refs from the accessibility tree. Chrome uses Anthropic's computer use tool with pixel-level coordinates plus an accessibility tree read_page tool.

Bot detection: Both Playwright tools use automated browser instances that may trigger bot detection. Chrome uses your real browser fingerprint, making detection unlikely.

Maturity: MCP has been stable since March 2025. Chrome entered open beta in December 2025. CLI is the newest at v0.1.1 (February 2026).

Use Playwright CLI when you need token-efficient browser control at scale. Best for coding agents generating tests, scripted multi-step workflows, CI/CD browser automation, token-sensitive production environments, and parallel browser sessions via named sessions.

Use Playwright MCP when you need the full browser toolkit with any MCP-compatible agent. Best for autonomous exploration and testing, cross-browser compatibility checks, complex interactions requiring 25+ tools, Docker-based CI pipelines, and teams already using MCP infrastructure.

Use Claude in Chrome when you need to interact with authenticated apps without credential juggling. Best for testing behind login walls, visual QA with live screenshots, multi-site authenticated workflows, developer build-then-verify loops, and recording interaction GIFs for documentation.

The most powerful setup is not choosing one tool — it is mapping each tool to the development phase where it excels.

Phase 1 — Development: Use Claude in Chrome. Build a feature in your editor. Claude opens localhost in your real Chrome, takes a screenshot, checks console errors, verifies visual output. No separate browser needed — it uses your dev environment exactly as you see it.

Phase 2 — Test Generation: Use Playwright CLI. Generate E2E test scripts from the working feature. CLI's token efficiency means you can run dozens of snapshot-and-assert cycles without burning context. Named sessions let the agent test multiple pages in parallel.

Phase 3 — Cross-Browser QA: Use Playwright MCP. Run the generated tests across Chromium, Firefox, and WebKit. MCP's full tool suite handles complex assertions, network mocking, and device emulation. Docker support makes this CI-ready.

Phase 4 — Authenticated Flows: Use Claude in Chrome again. For flows that require real login sessions (OAuth, SSO, third-party integrations), Chrome bypasses the authentication problem entirely. Test the full user journey with your actual credentials.

Playwright CLI: Install globally with npm install -g @playwright/cli@latest. Run playwright-cli install --skills to add agent skills for Claude Code or GitHub Copilot. Use playwright-cli open --headed to launch a visible browser, playwright-cli goto to navigate, and playwright-cli snapshot to capture page state.

Playwright MCP: Add to Claude Code with claude mcp add --transport stdio playwright -- npx @playwright/mcp@latest. For Claude Desktop, add the server configuration to claude_desktop_config.json. Common flags include --browser (chrome, firefox, webkit), --headless, --viewport-size, --user-data-dir for persistent profiles, and --save-trace for debugging.

Claude in Chrome: Install the Claude for Chrome extension from the Chrome Web Store. Start Claude Code with claude --chrome or run /chrome within an existing session. The extension auto-registers as an MCP server, making all browser tools available immediately. Run /mcp and select claude-in-chrome to verify the tool list.

Playwright CLI runs isolated browser sessions by default with no persistent state unless the --persistent flag is used. File access is scoped to the working directory. It shares the PLAYWRIGHT_MCP_ environment variable namespace with MCP, which could cause confusion if both are configured simultaneously.

Playwright MCP is not a security boundary — origin filtering does not prevent redirects. File access is restricted to workspace roots by default. It supports --isolated mode for ephemeral profiles. All automation is visible in tool call logs.

Claude in Chrome performs domain-level permission checks before mutating actions, verifying that the tab's current URL still matches the originally permitted domain. However, Anthropic's own research showed an 11.2% prompt injection success rate even with safety mitigations. The extension has full access to your logged-in sessions — powerful but risky. All actions are visible in real-time in your browser window.

The browser automation landscape for AI agents is maturing fast. Microsoft has signaled that Playwright CLI will become the preferred interface for coding agents, while MCP remains the standard for autonomous, exploratory automation. Claude in Chrome occupies a unique niche — the only tool that works with your real browser state.

The winners will be teams that use all three strategically: CLI for token-efficient scripted tasks, MCP for cross-browser CI/CD, and Claude in Chrome for authenticated visual QA. As these tools mature through 2026, expect tighter integrations, lower token costs, and eventually convergence toward a unified agent-browser protocol.

For now, the practical advice is simple: start with the tool closest to your workflow. If you are already in Claude Code, try Claude in Chrome first. If you are building CI pipelines, start with Playwright MCP. If token costs are your bottleneck, move to Playwright CLI. Then expand as your needs grow.