Claude in Chrome MCPとCowork、CLIからブラウザを操る2つの方法——実測で分かった使い分け A hands-on comparison of two ways to control a browser from Claude Code CLI—Cowork and the…

The tooling for controlling a browser directly from Claude Code's CLI is maturing quickly, and developers now have two distinct options to choose from: Cowork, which routes through Anthropic's Claude Desktop application, and the Claude in Chrome MCP, a Chrome extension that exposes browser control via the Model Context Protocol.

Cowork leverages Claude Desktop's agent integration layer, giving it awareness of the broader desktop environment. This makes it well-suited for tasks that combine browser activity with other desktop context—think confirmation steps that involve screenshots, or workflows that jump between a web interface and a local file. The trade-off is that Claude Desktop must be running as a prerequisite, which adds some overhead to the setup.

Claude in Chrome MCP takes a lighter approach. Installed as a Chrome extension, it runs an MCP server locally and accepts tool calls from Claude Code over the protocol directly. This means tab management, DOM inspection, JavaScript execution, and form interaction are all available from the CLI without launching any additional desktop software beyond Chrome itself. In hands-on testing, this approach reportedly yields faster round-trip times for lightweight tasks and offers considerable flexibility for scripting.

Understanding where each tool shines requires a bit of context about MCP itself. The Model Context Protocol is an open standard introduced by Anthropic in late 2024 to give LLM agents a consistent interface for connecting to external tools, resources, and data sources. Since then, the ecosystem has grown rapidly, with MCP servers covering everything from file systems and databases to browser automation and third-party APIs. Claude in Chrome MCP sits squarely in this wave, but its defining characteristic is that it operates inside a real, live Chrome session—not a headless instance. That means the automation inherits whatever state the browser already holds: cookies, logged-in sessions, installed extensions, and so on.

This is a meaningful distinction from tools like Playwright or Puppeteer, which typically spin up isolated browser contexts. For tasks that require authentic user state—automating a SaaS dashboard you're already logged into, for instance—the live-session approach removes a whole class of authentication friction.

The broader competitive landscape is worth noting. Microsoft has been deepening the integration between Edge and Copilot for agentic browsing scenarios, while open-source projects like browser-use and Stagehand are gaining traction as framework-agnostic alternatives. Claude in Chrome MCP has the advantage of tight integration with Anthropic's toolchain, though its Chrome-only scope is a limitation compared to cross-browser solutions.

Based on the hands-on findings reported in the article, a reasonable heuristic emerges: reach for Claude in Chrome MCP when you want fast, scriptable browser automation from the CLI and your workflow lives inside Chrome; lean on Cowork when the task is more complex, involves desktop-wide context, or benefits from Claude Desktop's richer agent orchestration. Both tools are still evolving, and the gap between them could narrow—or widen—as Anthropic iterates on each. For now, knowing the strengths of each approach is the most practical edge a developer can have.