A← Index

Playwright MCP vs Chrome DevTools MCP vs bare: which browser should an AI agent use?

Three MCP servers that let a coding agent open, read and test web pages.

IThe short answer

For most agent web testing, use Playwright MCP. It reads pages as accessibility snapshots, so the model does not need screenshots. It drives Chrome, Firefox, WebKit and Edge.

Use Chrome DevTools MCP when the agent must debug a page in Chrome. It records performance traces and reads network requests and console messages with source-mapped stack traces.

bare-mcp is my own option. It drives bare, a browser engine I write from scratch in Rust. It is a learning project. Use it only when you want journeys as step files and an engine you can read end to end.

IISide by side

Playwright MCPChrome DevTools MCPbare-mcp
EnginePlaywright: Chromium, Firefox, WebKitChrome and Chrome for Testing, through Puppeteerbare, written from scratch in Rust
How the agent reads a pageAccessibility snapshots; vision tools are opt-in with --caps=visionDevTools data and screenshotsText and snapshot first; screenshots on request
Debugging toolsConsole and network by default; storage, tracing and video as opt-in capabilitiesPerformance traces, network, console with source mapsConsole, network, storage, diagnose
WindowHeaded by default; --headlessLive ChromeHeadless, or a window on Linux (Wayland)
Maintained byMicrosoftThe Chrome DevTools teamArjun P; a learning project, not for everyday browsing
LicenseApache-2.0Apache-2.0MIT or Apache-2.0
Startnpx @playwright/mcp@latestnpx -y chrome-devtools-mcp@latestcargo run -p bare-mcp

From each project's README, checked on 12 November 2026.

IIIPlaywright MCP

Microsoft's server uses Playwright's accessibility tree, not pixels. A text model can click and type by element reference. Coordinate tools are available when you turn on vision.

It needs Node.js 18 or newer. Its README states that it is not a security boundary, so run it only where the agent may act.

IVChrome DevTools MCP

The Chrome DevTools team ships this server for coding agents that must inspect a live Chrome. Its strength is diagnosis: a trace with performance insights, the network log and console errors with source-mapped stacks.

Two defaults matter. Usage statistics are on, and --no-usage-statistics turns them off. Trace URLs can go to the Chrome UX Report API for field data, and --no-performance-crux turns that off.

Vbare-mcp

bare has its own networking, TLS, HTML and CSS parsers, layout, paint and a JavaScript interpreter. bare-mcp exposes journeys to an agent: open, find, act, read, snapshot, screenshot, console, network and storage, plus run_steps for a whole step file.

Input goes through the same path as a person, so a passing journey means the page works in the window. On 10 November 2026, bare passed 48 journeys on public sites. Layout and JavaScript are still in progress, and many sites fail. When a site fails, the agent files an issue on mrwick1/bare.

I built bare, so treat this section as the author's view.

VIQuestions

What is the best MCP browser for AI agents?

For general web testing, Playwright MCP is the common choice: it reads accessibility snapshots, supports Chrome, Firefox, WebKit and Edge, and needs no vision model. For debugging performance, network or console errors in Chrome, use Chrome DevTools MCP.

What is the difference between Playwright MCP and Chrome DevTools MCP?

Playwright MCP automates pages across several engines through accessibility snapshots. Chrome DevTools MCP controls and inspects a live Chrome, with performance traces, network requests and source-mapped console messages.

Does Playwright MCP need screenshots?

No. It works from structured accessibility snapshots. Screenshot and coordinate tools exist, and the coordinate tools are opt-in with --caps=vision.

What is bare-mcp?

bare-mcp is an MCP server for bare, a web browser written from scratch in Rust by Arjun P. Agents use it to run end-to-end journeys, headless or in a window. It is a learning project and is not safe for everyday browsing.

SignedArjun