agentic-playwright-mcp

agentic-playwright-mcp lets multiple agents control one Chrome instance without sending commands to the wrong tab. Each browser tool accepts a targetId. Agents can also use a groupId to list and control only their own tabs.

This package is an independent fork of Playwright MCP. It is not a Microsoft package.

Run the MCP server

Use this command as the server command in your MCP client:

npx --yes agentic-playwright-mcp

For clients that use JSON configuration, add:

{
  "mcpServers": {
    "agentic-playwright-mcp": {
      "command": "npx",
      "args": ["--yes", "agentic-playwright-mcp"]
    }
  }
}

The command starts an MCP server over standard input and output. It waits for requests from an MCP client, so running it directly does not open a prompt or web page.

Chrome starts when the client calls the first browser tool. The server uses one headed Chrome instance and stores its profile in ~/.agentic-playwright-mcp.

To restart Chrome if it exits, add --keep-alive to the arguments:

"args": ["--yes", "agentic-playwright-mcp", "--keep-alive"]

To connect to an existing browser, start Chrome with remote debugging and add --cdp-endpoint:

"args": [
  "--yes",
  "agentic-playwright-mcp",
  "--cdp-endpoint",
  "http://localhost:9222"
]

Related MCP server: SpidexLab's Multi-Tenant Browser MCP

Install the CLI

Install the package globally if you do not want the MCP client to run it through npx:

npm install --global agentic-playwright-mcp
apmcp --help

The package installs these commands:

• agentic-playwright-mcp and apmcp start the MCP server.

• agentic-playwright-mcp-daemon and apmcp-daemon manage the Chrome daemon.

Keep agent tabs separate

1. Call browser_tab_group with action: "create". Save the returned groupId.

2. Call browser_tabs with action: "new" and the groupId. Save the returned targetId.

3. Pass the same targetId and groupId to later browser calls.

Selecting a tab does not raise Chrome. If the window is minimized or behind another app, it stays there. Tools keep using targetId, so the tab does not have to be in front.

Cookies, local storage, and sign-in state remain shared because all groups use the same Chrome profile. A person can complete an interactive sign-in, then the agent can continue on the same targetId.

The package includes a companion extension and loads it through CDP when Chrome starts. When Chrome allows the extension, group listings include native groups that a person made in the tab strip. If browser policy blocks the extension, MCP group isolation still works.

Security

browser_run_code_unsafe runs JavaScript with access to the selected Playwright page. Only let trusted clients call this tool.

The Chrome DevTools Protocol endpoint listens on localhost:9223 by default. Do not expose this port to other machines.

Documentation

• Share one browser between agents

• Tool reference

• CLI reference

• Chrome extension notes

• Test suite

Develop from source

git clone https://github.com/TechyAditya/agentic-playwright-mcp.git
cd agentic-playwright-mcp
npm ci
npm run build
npm test
npm run test:e2e

Differences from Playwright MCP

Use Playwright MCP when one agent owns the browser. Use this package when agents must share a signed-in browser without sharing tab control.

License and attribution

This project includes work by Patrick Menlove and code from OpenClaw. See the license file for attribution and license terms.