gemini-skill drives Gemini's web app with a stealth plugin and your cookies
Gemini drawing MCP and skill via browser, usable with OpenClaw or any agent that supports MCP.
At a glance
- What is it?
- An MCP server that controls the Gemini web interface over the Chrome DevTools Protocol, with a daemon that keeps your logged-in browser alive for thirty minutes after the last call. It also carries a second, unrelated backend and two sponsored sections.
- Who is it for?
- Start with the terms of service, because that is the part this project cannot help you with. It drives Google's own web interface with a stealth plugin whose documented job is hiding the webdriver flag and emulating a browser fingerprint so the site's automation detection does not fire, and it operates inside an account you are logged into rather than through any API Google offers.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 9 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- WJZ-P/gemini-skill834 ★ · +2
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| WJZ-P/gemini-skill | 832 | 834 |
2 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The anti-detection plugin is a named design goal
The architecture section lists four core ideas, and one of them is evasion stated in so many words: a stealth mode that uses puppeteer-extra-plugin-stealth to get past the site's detection. The configuration section is more specific about what that means. The project's own browser instance is described as providing full anti-scraping protection by hiding the webdriver marker and simulating a real browser fingerprint, so as to better avoid the website's automation detection, and that is given as the reason to keep the default port rather than reuse another tool's browser. The comparison is drawn explicitly against reusing OpenClaw's session, which the page says has no stealth plugin and is therefore weaker on detection. That is not a hidden implementation detail; it is the reason the page recommends a particular configuration. Whether automating a consumer web front end this way is permitted is a question for the service's own terms, and nothing in this repository answers it.
Your Google session is persisted, and the browser outlives the call
Gemini needs a login, so the browser must already be signed into a Google account before anything works, and the profile that holds those cookies is kept on disk. The configuration is in a table with five browser variables. BROWSER_USER_DATA_DIR is the user data directory that stores login state and cookies, and if you leave it unset it resolves automatically, first to a dot-directory in your home folder and then to the browser's default directory. BROWSER_HEADLESS defaults to false, with the advice to keep it off the first time so you can log in. BROWSER_PATH auto-detects Chrome, Edge or Chromium rather than shipping a browser, which is why the dependency list uses puppeteer-core. The lifetime is the other half: the browser process is owned by a separate daemon, it does not close when an MCP call returns, and DAEMON_TTL_MS defaults to 1800000, thirty minutes of inactivity, after which the browser closes and the daemon exits and the next call starts it again. DAEMON_PORT is 40225, and the table gives no host or authentication setting for it.
MIT in the metadata, ISC in the package file
The licence is stated twice and the two statements are different families. The licence carried in the repository's own metadata is MIT, and a LICENSE file sits at the root next to README.md and README.en.md. The package manifest says something else:
"name": "gemini-skill",
"version": "1.1.1",
"type": "module",
"main": "src/index.js",
"author": "",
"license": "ISC",The author field is empty, so the manifest records no author either. For a package that people install into an agent runtime, the manifest is the artefact that travels, so this is the copy that ends up in tooling. Neither file is annotated as applying to part of the tree, and nothing in the repository says which one governs. The dependency set is small and tells you what the thing is: the Model Context Protocol SDK, puppeteer-core, puppeteer-extra, puppeteer-extra-plugin-stealth and sharp, which is the image library the post-processing path uses. Five scripts are defined, and three of them are demos named demo, demo2 and demo3.
A .env file is committed beside the .env.example template
The root of the repository contains both a .env file and a .env.example. The configuration section says every setting goes through environment variables or a .env file, that the project root already ships a template you can edit, and that the precedence order is process.env first, then .env.development, then .env, then the code defaults. The template itself is well behaved: it lists every variable with a bilingual comment and leaves the sensitive fields empty, with a note telling you to put the real key in .env.development or the local environment. And .env.development is described as not tracked by git and therefore suitable for local private configuration such as a browser path. That guidance only holds if the committed .env at the root is not where the real values ended up. The root also carries ATLAS_CLOUD_PROVIDER_REVIEW.md, a review document for the second backend discussed below, along with a SKILL.md, a references/ directory and a markdown/ directory.
The selectors are a running commentary on Google's interface
The most useful thing in the README is its update log, because it documents exactly how this project breaks. Version 1.1.0, dated 2026-05, adapted to the newest Gemini interface: the plus panel button, the send button, the download-full-size button implemented as a custom element, and the attachment-loaded detection, each named with its attribute or class, plus compatibility with a new wrapper element around the icon buttons. Every selector kept a fallback for the old interface so both versions work. Version 1.1.1 renamed the model enumeration from pro, quick and think to pro, flash and flash-lite, moved the current-model read to a different element's innerText, and changed model switching to iterate menu items and match label text for i18n safety. It also added thinking-depth switching and a new MCP tool for it. That is a project whose maintenance cycle is set by another company's front end. Note also that the version numbers live in a README section and in the manifest; the repository has no GitHub releases, so v1.1.1 is not a tag you can check out.
Two unrelated backends in one repository
The README carries a sponsored section for Atlas Cloud, described as a full-modal AI inference platform offering a unified API across more than 300 models, with a link carrying campaign parameters. Under it sits a second implementation: a minimal-intrusion provider that does not change the existing Gemini browser automation path, adds an OpenAI-compatible Atlas Cloud API provider, and supports model enumeration, plain chat and streaming chat verification. The stated purpose is to let the project serve as both a Gemini automation skill and a MaaS provider example. The configuration for it is a second block of variables: ATLAS_BASE_URL defaulting to the vendor's OpenAI-compatible endpoint, ATLAS_API_KEY with an empty default and a recommendation to put it in .env.development, ATLAS_MODEL defaulting to openai/gpt-4o-mini, and a request timeout of 60000 milliseconds. A third section points at Bloome, an agent IM platform, again behind a link carrying a referral parameter. Between the two, a good part of this README is vendor promotion rather than documentation.
The debug port is designed to be shared with another tool
The browser variable table has a note that is easy to miss: BROWSER_DEBUG_PORT defaults to 40821, and multiple skills, with douyin-upload-mcp-skill given as the example, sharing that port share one browser instance. The configuration section then makes the same sharing explicit for a named third-party tool. OpenClaw's default CDP port is 18800, and you are told you can set BROWSER_DEBUG_PORT to that value to reuse a browser session OpenClaw has already started. The page attaches a caveat and then a recommendation: that session has no stealth plugin integrated and is weaker on detection, so unless you have a specific reason, keep the default 40821 and let the project manage its own browser. The practical consequence is that this server will attach to whatever browser already listens on the port you name, running with whatever profile and login state that browser holds. Two projects, one authenticated Google session, one debugging port, and no isolation described between them.
Downloaded images come back with the Gemini watermark removed
One row of the feature table says automatic watermark removal: images that are downloaded have the Gemini watermark stripped. The image library in the dependency list is what does it, and the extraction row above it mentions base64 output and full-size download over CDP, using the custom element that version 1.1.0 had to adapt. So the pipeline is: drive the page, wait for the attachment to finish loading, take the full-resolution image rather than the preview, and post-process it before it lands in OUTPUT_DIR, which defaults to ./gemini-image. Whether a generated image should carry the mark of the service that generated it is a question the project does not address anywhere on this page, and it is the kind of detail that matters if the output is going anywhere public. One more boundary is worth stating plainly: the architecture diagram itself stops partway through, ending on an unclosed box-drawing frame, though the layering underneath is described in one line, from the MCP protocol layer through the operations layer and the connection layer to process management in the daemon.
Editorial conclusion
Start with the terms of service, because that is the part this project cannot help you with. It drives Google's own web interface with a stealth plugin whose documented job is hiding the webdriver flag and emulating a browser fingerprint so the site's automation detection does not fire, and it operates inside an account you are logged into rather than through any API Google offers. If automating your own account that way conflicts with how Google describes acceptable use, no amount of the project's own documentation settles it. Then weigh the rest on its merits. What you get is a working MCP surface for image generation, chat, upload and extraction, with watermark removal on downloads, and a selector layer that has to be patched every time Google changes its interface, which the project's own changelog documents happening twice in one month. Two more things to check before you rely on it: the licence is stated two different ways in two files, and a .env file sits in the repository root next to the example template while the documentation tells you to keep private keys somewhere git does not track.
Frequently asked questions
What does gemini-skill need installed before it will run?
Node.js 18 or later, a Chrome, Edge or Chromium install or a BROWSER_PATH pointing at one, and a browser already logged into a Google account since Gemini needs a login. Installation is a git clone, then cd gemini-skill, then npm install.
Where does gemini-skill keep my browser login?
In BROWSER_USER_DATA_DIR, which auto-resolves first to a dot-directory in your home folder and otherwise to the browser's default directory, and it holds the login state and cookies. The page says private keys belong in .env.development, which git does not track.
How long does the gemini-skill browser stay open after a call?
The daemon keeps it alive and closes it after DAEMON_TTL_MS of inactivity, which defaults to 1800000, thirty minutes. The next call starts the daemon again. The daemon itself listens on DAEMON_PORT, default 40225.
Which licence is gemini-skill under?
The licence carried in the repository's own metadata is MIT and a LICENSE file is at the root, while package.json declares ISC and leaves the author field empty. Nothing in the tree says which statement governs.
What is the current version of gemini-skill?
package.json says 1.1.1 and the README's latest-updates section heads v1.1.1 for 2026-05. The repository has no GitHub releases, so those version numbers exist in the manifest and the README rather than as tags.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/wjz-p-gemini-skill)Star history chart for your README
[](https://hysenlabs.com/projects/wjz-p-gemini-skill)Was this useful?