Repository navigation
stealth mcp servers compared
Most browser MCP servers drive a stock automation build. A small, young cluster does not: wrappers around Camoufox, nodriver and Patchright, plus this project. This page is what separates them, and it starts with the disclosure, because one of the entries is ours and because two of the others make claims we could not check.
The wrapper is thin. What you are choosing is the engine.
Camoufox is a Firefox fork with anti-fingerprinting work done in the C++ source and driven over Juggler rather than CDP. Around 11.1k stars when the 2026 stealth-browser survey read it in August 2026, and actively pushed. The significant property is where the changes live: in the engine, so a page cannot observe the patching the way it can observe a script that rewrites properties after load.
nodriver is the successor line to undetected-chromedriver: Chromium, driven over CDP directly with no WebDriver layer. Around 4.6k stars in the same survey, with its last push in May 2026, which is worth knowing before you build on it.
Patchright is a patched Playwright driver: same API you know, with the best-known Playwright tells removed. Around 4.1k stars, pushed in August 2026. It is the smallest change of the three and the easiest to adopt, because your existing code keeps working.
This project's engine is invisible-playwright, a Firefox patched at the C++ source with the identity derived from a seed, so two runs with the same seed present the same browser.
Camoufox and this engine are the same family of idea and are genuine alternatives to each other. nodriver and Patchright are the Chromium answer to the same question.
mcp-camoufox exposes Camoufox over MCP with a large tool surface, installed through npx. mcp-stealth-chrome wraps nodriver together with a TLS-spoofing HTTP client. patchright-mcp-lite is the opposite design: four tools, one job. invisible_playwright_mcp is ours, and the server ships inside the same package as the interface: the MCP server has the config block.
Tool count is the axis these wrappers advertise on and it is the least useful one. A hundred tools is a hundred descriptions in the model's context on every turn, and a larger set to choose wrongly from. What matters is whether the model can read state without scripting, reach what selectors cannot, and fall back to JavaScript for the remainder. How the tools are shaped argues that in full.
Two of these projects advertise capabilities in their own repository text that we did not test and are not repeating as fact: a "proven" one-line bypass of a specific challenge product, and a vision-based solver for another. They may work. We have not run them, and a claim in a README is a claim.
We are explicit about this because the same restraint applies to us in the other direction: invisible_playwright_mcp does not solve captchas and does not promise non-detection. A patched engine changes what the browser looks like. It does not read a challenge for you, it does not repair a datacenter IP's reputation, and it does not slow down an agent that is clicking faster than a person could. Can an AI agent solve a captcha is the longer answer, and it is the same answer we would give about anyone's tool.
If you are evaluating any server in this category, the useful test is not reading claims. It is running the same page through two of them from the same IP, and this wiki's engine documents how to test that yourself.
- You are already on Playwright and want the obvious tells gone: Patchright, because your code does not change.
- You want Firefox with engine-level work and a mature project: Camoufox, or ours. Camoufox has the larger community; ours has seed-derived identity and the agent interface in the same package.
- You are committed to Chromium and want no WebDriver layer: nodriver, with the maintenance tempo noted.
- You want the smallest thing that works: patchright-mcp-lite.
- Your problem is not the browser: none of these. Read why an agent gets blocked and check the IP, the pacing and the rate limit first. That is the majority case, and this entire page is about the minority one.
What is a stealth browser MCP server? An MCP server whose browser is a modified engine rather than a stock automation build, so what a page can observe about the browser differs.
Is there a Camoufox MCP server? Yes, as a community wrapper. Camoufox itself is the engine, not the server.
Does a stealth MCP server get past Cloudflare? Not as a property of being a stealth MCP server. Cloudflare and Playwright MCP goes through what a challenge is actually reading.
Is Patchright the same as playwright-stealth? No. playwright-stealth patches the page from JavaScript after load; Patchright patches the driver. The engine wiki has the distinction in detail.
Which one is safest to depend on? Judge by push date and issue tempo, not by star count. The survey figures above are dated for that reason, and MCP servers on GitHub is the same judgement applied to any listing rather than to this one category.
Is a stealth server even the category I want? Only if the site recognising the browser is your actual complaint. How to choose among MCP servers puts this category next to the others first, which is the cheaper question to answer.
See also: Choosing an MCP server for browser automation, Playwright MCP alternatives, and open-source agentic browsers.
- Stealth browsers 2026: nodriver, Camoufox, Patchright, retrieved 2026-09-10, for star counts and last-push dates as of August 2026.
- RobithYusuf/mcp-camoufox and RobithYusuf/mcp-stealth-chrome, retrieved 2026-09-10, for the wrappers and their own claims.
- patchright-mcp-lite, retrieved 2026-09-10.
- invisible-playwright, this project's engine.
Written while maintaining one of the servers listed. The section naming the unverified claims applies to us first: the paragraph saying what invisible_playwright_mcp does not do is there so the rest of the page can be read as description rather than advertising.