Repository navigation
playwright mcp alternative
Microsoft's playwright-mcp is Apache-2.0, 36.9k stars when read on 2026-09-13, and maintained by the Playwright team. If you cannot name what it fails to do for you, there is no reason to replace it. This page is organised around the four complaints that are real, because three of them have a fix that is not a different server.
This page is about swapping one MCP server for another. If your question is whether to use MCP at all rather than a library, a function call or a plain script, that is a different question with a different answer: MCP alternatives.
The most common one, and it arrives as browser is already in use. A persistent profile can only be opened by one browser at a time, and the server defaults to a single fixed profile.
Fix without switching: --isolated for a throwaway profile per session, or
a distinct --user-data-dir per client. Both are documented options.
When a switch is warranted: if your client only lets you register a bare command with no arguments, you may not be able to pass either flag, and a server that has no shared profile to fight over saves you the fight. invisible_playwright_mcp is one: a server is one identity for its whole life, so two registered servers are two browsers by construction rather than two clients racing for one directory.
Every tool description is in the model's context on every turn, and a browsing session is a lot of turns. Reasonable complaint, wrong first move.
Fix without switching: trim the capability set. --caps is additive, so do
not enable vision, pdf and devtools if you are not using them. Do not register a
second browser server alongside it "just in case": the overlapping tool names
also make the model choose worse, not just cost more.
When a switch is warranted: if you genuinely need one job done, a four-tool server does it with a fraction of the surface. That is the design point of the smallest wrappers listed in stealth MCP servers compared.
The one case where the alternative really is a different server, because the cause is the engine and no flag reaches it. playwright-mcp drives a stock automation build; that is a specific set of observable choices and it travels with the server.
Before switching, attribute the failure. Most reports of "the agent gets detected" are machine and network facts that no server changes: a datacenter IP, no GPU, a container font set, or an agent clicking faster than a person can. Why an agent gets blocked works through how to tell which one you have. Switching servers over an IP problem costs you a day and fixes nothing.
If it really is the browser: the engine-modified servers are the row that differs. Camoufox, nodriver and Patchright wrappers, and this project. Stealth MCP servers compared is the comparison, including which of their claims we could not verify.
Sometimes the alternative to an MCP server is not an MCP server.
If you are running the same known flow every night, a model rediscovering it each time is slower and more expensive than a script. Playwright MCP vs the CLI has the split. If you are pulling structured data at volume, see MCP for web scraping: the working pattern is a model to establish the shape once and plain code for the repetition. If you want an autonomous program rather than a tool inside your assistant, you want an agent library - browser-use, Skyvern, Stagehand - and open-source agentic browsers lays out that layer.
Engine-modified MCP servers. Camoufox, nodriver and Patchright wrappers. Address complaint 3 only. Younger and smaller than Microsoft's, so check push dates.
invisible_playwright_mcp (ours, MIT). A Firefox patched at the C++ source, identity derived from a seed so a failing run replays exactly, one identity per server with a helper browser that shares nothing with it, and the agent interface in the same package. Addresses complaints 1 and 3. It does not make the model better, does not solve captchas, and covers Windows and Linux only.
Agent libraries. Not MCP at all. The right answer to complaint 4 when you wanted a program.
Staying, with flags. The right answer to complaints 1 and 2 most of the time, and the cheapest.
Is there a better Playwright MCP? Better at a named thing, yes. Better in general, no, and any page telling you otherwise is selling something.
What is the closest alternative to Microsoft's server? Functionally, the Camoufox and Patchright wrappers - the same shape with a different engine.
Can I run two MCP browser servers at once? Yes, and it costs context on every turn plus a harder choice for the model. Prefer one, configured properly.
Does an alternative server make my agent undetectable? No. Nobody can say that honestly. The engine changes what the browser looks like and nothing else.
Is playwright-mcp being maintained? Yes, actively, by the Playwright team. The alternatives are the ones where maintenance tempo is worth checking.
See also: Choosing an MCP server for browser automation, Playwright MCP best practices, and the MCP server.
-
microsoft/playwright-mcp, retrieved 2026-09-10: licence, stars,
--isolated,--user-data-dir,--caps. - The wrappers and engines surveyed in stealth MCP servers compared, with dates.
Written while maintaining an alternative to the tool this page is about. Three of the four complaints resolve without us, and they are listed first for that reason.