Modified Firefox browser that masks its fingerprint for web scraping and AI agents

Many websites try to spot automated browsers by reading their fingerprint: the small details a browser reveals about itself. Camoufox is a Firefox-based browser built to hide those signs, so scrapers and AI agents get blocked less often.

The GitHub topics point to Playwright, so it is meant to plug into that automation tooling. The README says it is intended for use with rotating proxies, preferably residential IPs, since your IP address is something the browser cannot hide.

The README text available here is mostly sponsor listings. Full documentation is at camoufox.com, and the project warns that it is under development and may not suit stable production use. Install steps and feature details were not in the text I saw.

Read more on interesting repos · GitHub

Playwright controls real web browsers from code. You write steps like open a page, click a link, check a heading, and Playwright runs them in Chromium, Firefox or WebKit with the same API. It works on Linux, macOS and Windows, headless or with a visible window.

It comes in several forms. Playwright Test is a test runner with auto-waiting, retrying assertions and a fresh browser context for every test. The library is for plain scripts such as screenshots, PDFs and scraping. A CLI and an MCP server let coding agents and LLM clients drive a browser.

For agents, the README says the CLI is more token-efficient than MCP because it avoids loading large tool schemas and accessibility trees into the model context. The MCP server instead gives agents a structured accessibility snapshot of the page, so no vision model or screenshots are needed. Playwright is also available for Python, .NET and Java.

Read more on interesting repos · GitHub

Most browser tools for agents, such as Browser-Use and Vercel's agent-browser, are libraries that drive some other browser. Your logins rarely carry over cleanly, and you and the agent end up competing for the same window. AI browsers like ChatGPT Atlas and Perplexity Comet ship their own built-in agent, and only that agent can drive them.

ego lite is one browser built for you and any agent you bring. Each agent gets a Space, an isolated workspace inside the same browser, and you keep browsing in yours. On first launch it can migrate your Chrome data, so the agent inherits your logins, cookies, extensions and bookmarks.

The agent connects through an ego-browser skill. Instead of calling one command at a time, it writes a JavaScript snippet that calls in-page tools (snapshot, fill, click, wait, navigate, capture), and the browser runs it in one pass. A Snapshot is the page turned into structured text a model can read. The README reports up to 2.5x faster runs than agent-browser on four tasks; those are the authors' own benchmarks.

Read more on interesting repos · GitHub

Many tasks only exist behind a website: filling forms, searching, signing in, booking. Browser Use gives a language model the ability to open a browser, look at pages, click and type, and keep going until the task is done. You describe the task in plain text and the agent works out the steps.

You can use it three ways. The Python library runs the open-source agent in your own code, with your choice of model. The CLI adds browser access to an existing agent such as Claude Code or Codex. The hosted cloud runs the agent and the browser for you. The library and CLI can use a local browser or a cloud one.

The README also describes BU2, a Browser Use model tuned for browser work, and an Anthropic SDK integration that exposes Browser Use as Claude's browser tools. The library is MIT-licensed and free, but model calls and cloud browsers cost money unless you run a local model through Ollama.

Read more on interesting repos · GitHub

Page Agent is an in-page GUI agent. You add one script, or install an npm package, and the web page gets an agent that can read its own interface and act on it. A user or developer types an instruction in natural language, and the agent carries it out on the page.

It works on text, not screenshots. The agent reads a text version of the page's DOM and sends it to a language model you configure, so it needs no vision model or special browser permissions. You supply the model name, base URL and API key. The README says most mainstream models work, including ones you run locally.

Many browser agents run from outside the page with Python and a headless browser. Page Agent runs inside the page and is aimed at client-side features, such as a copilot in your own product. The README says its DOM processing and prompt are derived from browser-use. An optional Chrome extension handles tasks across several tabs, and a beta MCP server lets outside agent clients drive the browser.

Read more on interesting repos · GitHub

Most agents either run on your whole computer or in someone else's cloud. peerd takes a third route: the agent lives in a browser extension and relies on the browser's own security boundaries, such as workers, origins, sandboxing and OPFS storage.

The main agent hands work to separate, keyless actors. Each page or compute environment gets one, with tools limited to that environment. Credentials, network rules, confirmations and audit stay with the extension. Code runs in sealed JavaScript notebooks, WASI tools, browser apps, or Linux virtual machines (Chromium only), none of which can reach your host operating system.

You bring your own model: a hosted provider with your key, a local model on localhost, or early WebGPU support. There is no peerd account, and the README says current builds send no product telemetry. Preview builds add browser-to-browser sharing over WebRTC; store packages leave that out.

Read more on interesting repos · GitHub

Language models read plain text better than raw HTML full of menus, scripts and footers. Crawl4AI loads a page in a real browser (Playwright), runs its JavaScript, and returns Markdown that keeps headings, tables and code blocks. The author says it began as a free alternative to a paid web-to-Markdown service.

You can use it three ways: as a Python library, as your own Docker server with a REST API and MCP support, or through the hosted Crawl4AI Cloud. The library and server are free under Apache-2.0. The cloud is pay-as-you-go and adds web search, answers, and handling of bot-protected pages.

Beyond Markdown, it can pull structured data using CSS or XPath schemas (no LLM needed) or an LLM into a typed JSON schema. It can also follow links across a site, manage sessions and proxies, and crawl many URLs at once.

Read more on interesting repos · GitHub

AI coding agents can edit code, but they often fail when asked to read the web. YouTube transcripts, Twitter search, Reddit threads and Xiaohongshu posts each need a different tool, login or workaround. Setting all of that up by hand takes a while.

Agent Reach is a capability layer, not a scraper of its own. For each platform it keeps an ordered list of backends, such as twitter-cli first and OpenCLI second. It tests them, picks the first one that really works, and installs it. The agent then calls those upstream tools directly, guided by a SKILL.md file that the project registers.

If a platform blocks a tool, the project moves to the next backend, so users don't have to change anything. The README gives one example: yt-dlp was blocked by Bilibili and was replaced with bili-cli. The `agent-reach doctor` command reports which channels work and which backend each one uses.

Read more on interesting repos · GitHub

Agents often run on a different machine from the one you browse on, such as a cloud VM or a homelab server. That means logging in twice, once on your Mac and once where the agent lives, for every site. agentcookie removes the second login.

A source process on your Mac watches Chrome's Cookies file, decrypts the entries with the macOS Keychain, filters them by a policy, and seals them with a key made during pairing. It sends them over HTTPS on your Tailscale network. A sink process on the Linux box receives them and injects them into the running Chrome through the Chrome DevTools Protocol. New tabs and browser contexts get the cookies too.

It is one-way and unattended, unlike tools built for merging browser profiles between two laptops that a person uses. The README states the limits: Google and Workspace sessions bound to device keys (DBSC) stay logged out unless someone signs in on the box. On Linux the default policy ships nothing until you write a blocklist or allowlist.

Read more on interesting repos · GitHub

TARS is a multimodal agent stack. It lets an AI model look at a screen, decide what to click or type, and carry out a task such as booking a hotel or drawing a chart. The repo ships two projects: Agent TARS and UI-TARS Desktop.

Agent TARS runs from a command line or a Web UI. It can control a browser by reading the page structure (the DOM), by looking at screenshots (a GUI agent), or by mixing both. Its core is built on MCP, so you can attach MCP servers to give it extra tools. It works with model providers such as Volcengine and Anthropic.

UI-TARS Desktop is a native app that runs a GUI agent on your own computer, powered by the UI-TARS model and Seed-1.5-VL/1.6 models. It has local and remote operators for computers and browsers. The README points to Midscene for browser-only use.

Read more on interesting repos · GitHub

Many sites check whether a visitor is a script by reading browser fingerprints: canvas, WebGL, GPU, screen size, the webdriver flag and more. Stock Playwright or Puppeteer gives these away, so scrapers and agents get blocked or hit CAPTCHAs.

CloakBrowser takes a different route from tools like playwright-stealth, undetected-chromedriver and puppeteer-extra, which inject JavaScript or change flags. The README says it changes Chromium's own source code, with 87 patches, and ships the result as a binary. The wrapper downloads that binary (about 200MB) on first run and hands you a normal Playwright-style browser object.

It does not solve CAPTCHAs and has no built-in proxy rotation. You bring your own proxies, ideally residential. The README says the newest build is free with a GitHub sign-in for one concurrent session, and paid Pro plans add more sessions. An older v146 build stays free on GitHub Releases. Those detection results are the project's own claims, so test them on your target sites.

Read more on interesting repos · GitHub

Scrapers break when a site changes its layout or blocks automated visitors. Scrapling tries to cover both problems in one library. You can fetch a single page or run a crawl over many pages with the same tools.

It has three parts. Fetchers download pages, including a stealth option that uses a headless browser and, per the README, handles Cloudflare Turnstile. The parser selects elements with CSS or XPath. If you save elements with auto_save, you can pass adaptive=True later to find them again after the page structure changes. A Spider class lets you write a parse method and run concurrent crawls with pause/resume, proxy rotation, and a request pace that slows down when a site starts blocking.

The README also lists a command-line interface, an MCP server and an agent skill, so AI tools can use it for scraping. It doesn't compare itself with other tools by name. It suits people who would otherwise stitch together an HTTP client, a browser driver and a parser by hand.

Read more on interesting repos · GitHub

Agents often need to use websites: log in, fill a form, read a page, check how something looks. agent-browser gives them a simple set of shell commands for that, so any agent that can run a terminal command can control Chrome.

The usual loop is short. You open a URL, then run snapshot to get the page's accessibility tree, where each element has a ref such as @e2. The agent then clicks, fills or reads text using those refs. It also accepts ordinary CSS selectors and semantic finders like role, label and text.

It is a native Rust binary installed through npm, Homebrew or Cargo. The README says the daemon does not need Playwright or Node.js, and that it detects existing Chrome, Brave, Playwright and Puppeteer installs. It also has a read command that fetches pages as agent-friendly text without launching Chrome, and experimental WebMCP support.

Read more on interesting repos · GitHub

Language models can't use a web page on their own. This project is a Model Context Protocol (MCP) server that gives any MCP-aware AI client a set of browser tools, built on Playwright, Microsoft's browser automation library.

Instead of sending screenshots, it gives the model an accessibility snapshot of the page: a structured text description of buttons, links, fields and other elements. The model works from that text, so it needs no vision model, and its actions are less ambiguous than guessing from pixels.

The README compares it with Playwright CLI plus SKILLS, a separate project. It says the CLI uses fewer tokens and suits coding agents with limited context. The MCP server fits longer loops that need a persistent browser session and rich page detail, such as exploratory automation or self-healing tests, where token cost matters less.

Read more on interesting repos · GitHub