Skip to content

playwright mcp blocked

github-actions[bot] edited this page Nov 13, 2026 · 4 revisions

Playwright MCP session getting blocked

"Blocked" covers four failures that look the same from inside a chat window and need completely different responses. Getting the attribution right is most of the work, because three of the four are not detection at all.

Tell them apart first

The snapshot is empty or missing what you can see in a browser. Not blocked. The page rendered its content after the snapshot was taken, or the content is inside an iframe the snapshot did not descend into, or it is drawn on a canvas that has no accessibility representation. Ask for the snapshot again after the page settles, or take a screenshot and compare: if the screenshot has the content and the tree does not, this is your case.

A page that says something about verifying you are human. A challenge. This is a judgement made mostly before your page interaction, on the address and the connection. Cloudflare and Playwright MCP covers what one is actually reading.

It worked for a while and then stopped. A rate limit, almost always. The signature is that it is time-based rather than immediate, and that waiting fixes it. No browser setting reaches this. Slow down.

It fails immediately, on the first request, from this machine only. Now you have a candidate for either the address or the browser. Those two are separable and the next section separates them.

Address or browser

Run the same page from the same machine in an ordinary browser you use by hand.

  • The hand browser is fine, the agent is not: the browser is a candidate. Same address, different result.
  • Both are refused: the address. Nothing about your server or engine matters yet, and switching servers will waste a day.

Then run the agent from a different network, a phone hotspot is enough.

  • Different network, works: confirmed address.
  • Different network, still refused: now it is the browser or the pacing.

This costs ten minutes and is the difference between fixing the problem and switching tools at random. Why an agent gets blocked has the longer version of the same reasoning.

What each cause actually responds to

Address. A proxy, and the type matters: Playwright MCP with a proxy, including the three leaks that a proxy does not close by itself.

Pacing. Fewer actions per minute, and pauses that are not identical. An agent under a model is naturally slower than a script, which helps, but a model told to check forty pages will do it in a rhythm no person has.

The browser. This is the one where the server choice matters, and only here. A stock automation build is a specific set of observable choices that no flag changes. Engine-modified alternatives exist and stealth MCP servers compared lists them, along with the claims among them that we could not verify.

The snapshot. Not blocked. Read again, or screenshot.

What none of it responds to

A challenge you are meant to solve interactively stays a challenge. This project does not solve captchas and does not claim non-detection: can an AI agent solve a captcha is the direct answer, and it applies to every tool in this category regardless of what its README says.

If a site is refusing an agent deliberately and consistently, that is a decision the site made. The useful question at that point is whether the data is available another way, not which server to try next.

Short answers to the questions that lead here

Why does Playwright MCP get blocked and my normal browser does not? Usually not the same test. Your normal browser has history, a profile and a session; the agent's is fresh. Sometimes it is the address. Occasionally it is the automation build.

Does --isolated make blocking worse? It can. A profile with no history is one more thing that is unusual, though far less decisive than the address.

Will a different MCP server fix it? Only if the browser was the cause, which is the least common of the four. Attribute first.

Is a headless browser more likely to be blocked? Headless is detectable in its own right, and running headed costs you nothing on a desktop. Try headed before concluding anything.

Why does it work locally but fail on the server? The classic case, and it is the machine and the address rather than the code: no GPU, a container font set, a datacenter address.

See also: when the agent gets blocked, browser-use getting blocked, and Claude computer use detected as a bot.

Sources


Written while maintaining a server whose differentiator is the browser. The page puts the browser fourth out of four causes, because that is where it belongs by frequency.

Alternatives and Comparisons

When the Agent Gets Blocked

Using the Agent

Clone this wiki locally