Skip to content

Playwright MCP 0.0.84 adds isolated contexts per tab

Updated Playwright MCP

Summary

Playwright MCP v0.0.84 (October 31, 2026) adds named isolated browser contexts, URL parameter redaction, capped find results and network history.

Category
MCP servers
Date
Open source
Yes
License
Apache-2.0
MCP
Yes
Pricing
Not stated in the sources we cited.

Microsoft published Playwright MCP v0.0.84 on October 31, 2026, about five weeks after v0.0.83. npm moved @playwright/mcp latest to 0.0.84 the same afternoon. The headline is separate browser sessions inside one browser, so an agent can work with two accounts without them sharing cookies. Most of the other changes cut down on what lands in the model's context.

Release facts

FactDetail
ProductPlaywright MCP, npm @playwright/mcp
Versionv0.0.84
Release dateOctober 31, 2026 (GitHub release and npm)
Previous releasev0.0.83, September 28, 2026
LicenseApache-2.0
Runnpx @playwright/mcp@0.0.84 (Node.js 18 or newer)

New

  • Isolated contexts in one browser. browser_tabs with action: "new" takes isolatedContext: "<name>". Tabs with the same name share cookies and storage, and different names stay apart. Omit it, or pass "default", for the default context. Cookie and storage tools act on the current tab's context.
  • Redact URL parameters. --redact-url-params=access_token,code (config redactUrlParams, env PLAYWRIGHT_MCP_REDACT_URL_PARAMS) replaces those query and fragment values with <redacted> in tool responses. That covers tokens minted at runtime, which a secrets file can't list ahead of time. The notes say this is a convenience, not a security feature.
  • Capped find results. browser_find accepts maxResults and still reports the total match count.
  • Network history across navigations. browser_network_requests with all: true keeps up to 200 requests from earlier pages, and request indexes stay stable.
  • Locators in your language. browser_click and other element tools accept JavaScript, Python, Java and C# locators in target, alongside snapshot refs and CSS selectors.
  • --ignore-roots skips the roots/list request for clients that advertise roots but never answer it, which could stall the first tool call.
  • Cookie delete by domain and path in browser_cookie_delete (opt-in with --caps=storage).

Fixes worth knowing

  • HTTP transport: loopback hosts (localhost, 127.0.0.1, [::1]) are allowed for any bind address. Other hosts still need --allowed-hosts. With --isolated, each client now has its own idle timer.
  • Proxy: credentials in a --proxy-server URL are passed to the browser as proxy username and password.
  • Waiting: browser_wait_for can combine time, text and textGone. With a text condition, time is the timeout (capped at 30 seconds), not a delay.
  • Snapshots: unrendered text, inert elements and content outside the top modal <dialog> are left out. Click and fill wait for inert targets instead of misfiring.
  • Shutdown: a stdio disconnect closes the browser cleanly instead of force-killing it.

Questions

Should I upgrade? Yes if your agent logs into more than one account, or if tool responses leak OAuth code or access_token values into transcripts. The release notes list no breaking change.

Does redaction make it safe to give an agent real credentials? No. The release notes call --redact-url-params, like the secrets file, a convenience rather than a security feature. It only masks values in tool responses.

MCP or Playwright CLI? The README suggests Playwright CLI with skills for coding agents that are short on context, and MCP for long-running loops that need persistent browser state. See the Playwright MCP hub.

More: the MCP servers category, Browser Use for a Python browser agent, and this week's FastMCP 4.1.4.

Alternatives

Playwright MCP compared with its alternatives
ToolLicensePricing note
Playwright MCPThis launchMicrosoft's Apache-2.0 MCP server that lets LLM agents drive a browser through Playwright accessibility snapshots instead of screenshots.
Apache-2.0Apache-2.0 and free to run with npx. You pay your model provider for the agent that calls it.
Browser UseBrowser Use is an MIT-licensed Python library for AI browser agents that run locally with any LLM, plus the hosted Browser Use Cloud.
MITThe library is MIT and free. The README lists Browser Use Cloud stealth browsers from $0.02 per browser-hour.
TinyFishHosted web agent with an MCP server for search, browsing, and extraction, plus an MIT-licensed local proxy.
MITThe proxy repository is MIT. Hosted calls need a TinyFish account or API key. No dollar price is listed here.

Questions

What is Playwright MCP?

An MCP server from Microsoft, published on npm as @playwright/mcp, that gives an MCP client browser automation through Playwright. The README says it works on structured accessibility snapshots, so no vision model is needed.

How do I install Playwright MCP?

Add a server entry that runs npx @playwright/mcp@latest to your MCP client config. The README lists Node.js 18 or newer and clients such as VS Code, Cursor, Windsurf, Claude Desktop and Goose.

Should coding agents use Playwright MCP or Playwright CLI?

The README says coding agents may do better with Playwright CLI plus skills, because CLI calls use fewer tokens than large tool schemas and accessibility trees. It keeps MCP for long-running loops that need persistent browser state, such as exploratory automation or self-healing tests.

What is the latest release covered here?

v0.0.84, published October 31, 2026, which adds isolated browser contexts per tab name and URL parameter redaction.

Filed on Agent Radar. Claims in this note are limited to the sources above.