Let Claude Directly Operate Your Browser: Fill Forms, Post Content, Even Drag Slider Captchas

Most people still use AI in a "you ask, it answers" mode. But many truly annoying tasks aren't about "not knowing the answer" — it's about having to manually go through web pages step by step: filling out forms field by field, pasting curated content into some web editor to post, flipping through dozens of pages to copy information down, or even manually dragging a slider captcha when logging in. These are tasks AI should be best at helping you with, yet it can't reach your browser.

PandaNpc's Chrome plugin bridges this gap: It lets your Claude Code actually reach into your browser, click, fill, flip pages, and even drag captchas for you. It doesn't just take a screenshot for it to look at and tell you how to click — it does the clicking itself.

What It Is

In a nutshell: a Chrome extension that turns Claude into the "operator" of your browser.

The chain is simple:

你的 Claude Code  →  cc-relay 中转  →  浏览器插件  →  操作你当前看的网页

You say something like "help me post this content", "fill out this form", or "log into this website" in the conversation, and Claude invokes the plugin to execute actions on the tab you are currently looking at. The plugin acts as its "hands," and you can see everything it's doing in real time.

What It Can Do

The plugin exposes 43 browser operation tools to Claude, covering almost everything a person can do with a mouse and keyboard:

• See: Read page content, take screenshots, get the structure of a specific element, capture network request data behind the page.
• Click and Fill: Click buttons/links, fill input fields, checkboxes, dropdown selections, press keys (including combos like Ctrl/Enter), scroll, hover, drag.
• Navigate and Jump: Open/close/switch tabs, go forward/backward, navigate to URLs, wait for a specific element to load.
• Transfer: Directly upload an image (e.g., a cover photo) into a web upload box via URL, without the image bytes passing through the conversation.
• Handle Captchas: Identify slider gaps, drag sliders, click text/image captchas (detailed below).

The key isn't "how many tools there are," but that you don't need to remember them. You speak naturally, and Claude breaks it down into a sequence of actions to execute. Abstractly, the tasks it can help you with include:

• Handing over tedious and repetitive web operations in one sentence — filling forms, batch processing, moving information across pages;
• Pasting a curated piece of content into a web editor and publishing it in one go;
• Handling processes that require login or captcha solving.

You decide which website to act on and what to do.

Highlight: How It "Drags" Captchas

This is where the plugin truly shines, and where regular web scripts/Greasemonkey scripts fall short.

First, why ordinary automation fails at captchas. The browser stamps every event with a "truth seal" called isTrusted: events triggered by a real human (mouse clicks) have isTrusted=true; events generated by JavaScript code (element.click(), dispatchEvent) have isTrusted=false. Captcha and anti-scraping systems check this seal, and synthetic events are immediately exposed. So Greasemonkey scripts simulating clicks won't fool captchas.

This plugin takes a different path — Chrome DevTools Protocol (CDP) injects input events at the browser kernel level, producing events with isTrusted=true, making them indistinguishable from real human actions. This is the foundation for tools like real_click / real_drag / real_type.

Being "real" alone isn't enough; it must also drag like a human. The drag trajectory for sliders is specifically calculated as a human-like curve, not a linear instant move:

• Variable speed: starts slow, speeds up in the middle, then slows down near the gap (just like a human hand dragging something), not uniform speed;
• Arc movement: slightly curves upward in the middle then drops back, simulating a human hand's natural non-straight line;
• With jitter: each step adds pixel-level random micro-shakes;
• Pauses: a slight pause after pressing (start), another pause before releasing, and random millisecond intervals between steps.

The gap location is also identified by the plugin itself: it takes a screenshot of the captcha, converts it to grayscale, and calculates vertical edge "energy" column by column — the two vertical edges of the puzzle piece and the gap have the strongest energy, matching them to locate the gap. Claude gets the coordinates, calculates how far to drag, and drags along the human-like trajectory. After dragging, it takes a screenshot to verify, and if incorrect, it makes micro-adjustments.

• Puzzle slider: Identify gap → drag to it.
• "Drag to the rightmost" linear slider: Measure the handle and the right end of the track → drag all the way.
• Click text/image captchas: Screenshot to see the required click order and each character's position → click sequentially based on coordinates.

One more hidden difficulty: many websites' login boxes and captchas are embedded in cross-origin iframes. Regular scripts using selectors can't reach them. But this coordinate-based real input operates at the rendering layer, unaffected by same-origin policies, allowing direct operation inside cross-origin iframes.

Comparison:

Note: Captcha solving is intended to help you perform actions you are authorized to do (e.g., logging into your own account), not for mass cracking or attacking others' websites.

Safety: It Won't Do Anything Behind Your Back

Letting an AI operate your browser raises the biggest concern: "Will it do something malicious without my knowledge?" The plugin has multiple layers of safeguards:

• Every operation requires approval: Before Claude clicks or fills, a card pops up in the sidebar telling you "It wants to do X on page Y." You can Allow / Always Allow (this site, this operation) / Deny.
• High-security websites are blacklisted: Online banking, payment, email, and account login pages (ICBC / CCB / CMB, PayPal / Alipay / WeChat Pay, Gmail / QQ Mail / Outlook, Google / Microsoft login, etc.) are all on a blacklist — any operation is denied outright without even asking.
• Only affects your current tab: It won't dig into other tabs you have open in the background.
• One-click pause master switch: Press anytime to immediately halt all automation.
• Operation log: The last 100 operations (time, page, what was done, success/failure) are recorded and inspectable.
• Multi-browser isolation: Even if the same account is installed on multiple browsers, commands from each session only go to the browser it's bound to — no broadcasting, no misrouting.

In other words: What it can do depends on your approval card by card; what it has done is fully traceable.

Three Steps to Get Started

1. Install the Plugin: One-click install from Chrome Web Store 👉 PandaNpc — Remote Control for Claude Code

2. Pair with Your Claude Code: Install and run pandapaw on your development machine:

   bashCopy

   # macOS / Linux
   curl -fsSL https://cos.pandanpc.com/pandapaw/install.sh | sh
   # Windows (PowerShell)
   iwr -useb https://cos.pandanpc.com/pandapaw/install.ps1 | iex

   Then run pandapaw qr to generate a two-minute valid one-time pairing code, paste it into the plugin sidebar, and click "Connect."

3. Let It Work: Open the web page you want to operate on, and speak naturally in the conversation — "Help me fill this form," "Log in and handle the captcha yourself," "Post this content." It takes care of the rest, popping up approval cards when needed.

Summary

AI is most useful not when it thinks for you, but when it acts for you. This Chrome plugin transforms Claude from a "chatbox advisor" into an "operator inside your browser": 43 operation tools, ability to bypass slider and click captchas, and capability to work within cross-origin iframes, all paired with a comprehensive safety system of approvals, blacklist, pause, and logging.

If you want AI to truly work on web pages for you, install the plugin now and connect it to your Claude Code.