Playwright MCP vs Agent Chrome: Which Should You Use in 2026?
Use Playwright MCP if you want repeatable automation in Chromium, Firefox or WebKit, in Docker or CI, and a clean browser is what you want. Use Agent Chrome if the agent should work in your real, signed-in Chrome profiles and you want to see what it is doing.
Playwright MCP opens a browser of its own by default, so it starts without your logins. It also has an extension mode that connects to a running browser.
Side-by-side
| Playwright MCP | Agent Chrome | |
|---|---|---|
| Made by | Microsoft | Open source, operated by Circo, Inc. |
| License | Apache-2.0 | MIT |
| Browsers | Chrome, Firefox, WebKit, Edge | Chrome 120 or newer |
| Default profile | Its own persistent profile per workspace | Your own Chrome profile |
| Your existing tabs and logins | --extension mode, which needs the Playwright Extension |
Yes, by default |
| How the agent sees a page | Accessibility snapshots | Accessibility tree with ref_N handles, plus screenshots |
| Install | npx @playwright/mcp@latest in your MCP config |
curl -fsSL https://agentchrome.dev/install | sh, then load the unpacked extension |
| Video | WebM through opt-in devtools tools, with chapters and action overlays |
MP4, WebM and GIF |
| Docker and headless | Yes (headless Chromium in Docker) | Not documented |
| Several profiles at once | --profile-dir-name picks one profile in extension mode |
All paired profiles at once |
Setup
Playwright MCP needs Node.js 18 or newer. Add this to your client's config:
{ "mcpServers": { "playwright": { "command": "npx", "args": ["@playwright/mcp@latest"] } } }
For Claude Code: claude mcp add playwright npx @playwright/mcp@latest. The browser is headed by default, and --headless switches it off. You can also run it as a standalone HTTP server with --port 8931.
Agent Chrome:
curl -fsSL https://agentchrome.dev/install | sh
Then load the extension once per profile. See install.
Which agents it works with
Playwright documents Amp, Antigravity, Claude Code, Claude Desktop, Cline, Codex, Copilot, Cursor, Factory, Gemini CLI, Goose, Grok, Junie, Kiro, LM Studio, opencode, Qodo Gen, VS Code, Warp and Windsurf. Agent Chrome works with any MCP client, and the installer configures Claude Code, Codex, Gemini CLI, Antigravity, Grok, Muse, Cursor, opencode and Windsurf.
Real browser and logins
Playwright MCP keeps a persistent profile for each workspace, for example ~/Library/Caches/ms-playwright/mcp-{channel}-{workspace-hash} on macOS. Logins you make there persist, but it is not your daily profile. Only one browser instance can use a profile at a time, so parallel clients need --isolated or a separate --user-data-dir. --isolated keeps the profile in memory only.
With --extension, Playwright MCP connects to a running Edge or Chrome and reuses existing tabs and logged-in state. It needs the Playwright Extension.
Agent Chrome works in your real profile from the start. See logged-in websites.
Multiple profiles
In extension mode --profile-dir-name picks the profile, and it defaults to the last used profile that has the extension. Agent Chrome keeps every paired profile online together and lets each call choose with browser: "Work". See multiple Chrome profiles.
Visibility and control
Playwright's video tools can show an animated cursor and action overlays in recordings. The pages I checked do not describe an on-page overlay or a stop control.
Agent Chrome shows a pulsing blue border, a status pill with a Stop button and a named tab group per session. The popup has Pause, Resume and Stop all.
Recording
Playwright MCP's devtools capability adds browser_start_video and browser_stop_video. Files default to video-{timestamp}.webm at 25 fps. There are also browser_video_chapter cards and action overlays. Agent Chrome's record saves MP4 or WebM (default 15 fps) up to 10 minutes, with a natural cursor and no overlay unless you ask for cursor: "agent". See recording.
Security
Microsoft states "Playwright MCP is not a security boundary." Origin lists do not apply to redirects, file access is limited to workspace roots by default, and browser_run_code_unsafe is described as RCE-equivalent.
Agent Chrome stores keys only as SHA-256 hashes, expires pairing codes after 15 minutes and lets you revoke each profile. Because it controls your real sessions, pages can contain prompt injection. See security.
Pricing and license
Playwright MCP is Apache-2.0 and Agent Chrome is MIT. I found no price for either.
When to choose each
Choose Playwright MCP if:
- You need Firefox or WebKit.
- You run in Docker or CI and want a throwaway browser.
- You like snapshot-based automation and Playwright's test-generation tools.
Choose Agent Chrome if:
- The work is behind your logins and you do not want to set up a separate profile.
- You want to watch the agent, pause it and stop it.
- You want one setup for Claude Code, Codex and Gemini CLI.
See also the pillar guide, Chrome DevTools MCP and testing web apps with AI.
Sources
Frequently asked questions
Does Playwright MCP use my real Chrome?
Not by default. It opens a persistent profile of its own. The --extension option connects to a running Chrome or Edge and reuses its tabs and logins, and it needs the Playwright Extension.
Can Playwright MCP record video?
Yes, with the devtools capability enabled it can start and stop a WebM recording and add chapters. Agent Chrome records MP4, WebM and GIF without enabling an extra capability.
Is Agent Chrome a Playwright replacement?
No. Playwright MCP covers more browsers and runs in containers. Agent Chrome covers your real, logged-in Chrome. Many teams use both.
Which one is faster to set up?
Playwright MCP is one npx entry in your config. Agent Chrome is one command plus loading an unpacked extension once per profile.
Which has more tools?
Playwright documents opt-in groups for network routing, storage, PDF and test assertions. Agent Chrome focuses on tabs, input, reading, forms, debugging and recording. See the tool reference.
Try Agent Chrome
One command sets up every agent on your machine. Or add https://agentchrome.dev/mcp as a connector in Claude or ChatGPT.