Not a headless clone in a data centre — your Chrome, already signed into your accounts, driven by clicks, drags and keystrokes shaped like a person's. Ask in plain English; watch it happen in the window in front of you. Anything that publishes stops and waits for you.
No scripts, no selectors, no recording a macro. You type the sentence; Claude finds the buttons. These are real prompts — copy one.
Reply to the new comments on my last LinkedIn post — draft each one and I'll hit Post.
Grab my Stripe test keys and put them in .env
.env. No copy-paste,
no key sitting in your clipboard.
Walk through our signup flow and screenshot every step into ./docs
Drag the seat slider to 50 and tell me what it costs
Fill this vendor form from our company profile doc
Download last month's invoices from the billing portal
A normal automation script jumps a cursor to coordinates and fires a click event. Real people don't do that, and modern sites can tell. ChromeBoost sends the whole gesture through Chrome's own input pipeline — a curved approach, a little tremor before it lands, pressure on the way down. Drag is where the difference is easiest to see:
Menus that open on mouseenter, row buttons that only appear at
opacity: 0 until you're over them, tooltips, chart crosshairs.
There's nothing to click until something hovers — so ChromeBoost hovers, and waits.
Sliders, canvas and WebGL apps, map panning, sortable lists, resize handles, signature pads. Anything you'd move with a hand rather than tap with a finger.
Keystrokes go through Chrome's real input path, so React, TipTap, ProseMirror, CodeMirror, Monaco and Stripe's contenteditable fields all register them — the editors that silently swallow scripted input.
Here's the thing that breaks most browser agents. A page paints a cookie scrim, a modal
backdrop, or an invisible catch-all <div> over the button. The
click hits that instead — and the agent gets told it worked. It then argues with a
button it never actually pressed. ChromeBoost checks what's on top before it
clicks.
Blocked. The click never reaches the button — but the agent reports success and carries on.
Landed. Same real click, same trusted event — it just isn't aimed at a banner. The page is left exactly as it was found.
It looks for a spot on the target that's actually exposed. A button half-hidden behind a sticky footer gets clicked on the half you can still see.
If a fixed header or floating bar is in the way, it scrolls the page so the target moves clear — the same thing you'd do — before trying anything cleverer.
And when something genuinely can't be reached, it names the element in the way instead of claiming success. No more debugging a click that never happened.
An agent with your logins should not be able to publish in your name, move your money, or pass your 2FA on its own — and this one can't. When a step is irreversible, ChromeBoost fills in everything around it, highlights the final button, and hands control back to you.
Posts, comments and replies get drafted into the composer. The Post button is highlighted and waits for your actual click. Nothing goes out unread.
It never types a password or a one-time code. It brings you to the field, points at it, and pauses. Secrets on screen are redacted out of what Claude reads by default.
Checkout and card details are yours to confirm. Captchas aren't solved or faked — ChromeBoost reports one is there and steps back.
ChromeBoost is free and open source, and there's no signup or telemetry anywhere in it. If it saved you an afternoon, you can buy me a coffee.
While Claude drives, a small panel sits on the page so you always know a session is in control and which one. Most tools weld that badge to the top-right corner — exactly where sites put their account menu and cart. This one you can move. Drag it by the header, double-click to collapse, × to dismiss. Try it:
Drop it near an edge and it snaps flush; resize the window and it stays in view. Where you put it is remembered — next page, next session, same corner.
Collapse to a pill that just shows the session name, or dismiss it completely.
Alt+Shift+B brings it back on any page, and the toolbar popup has a
Show button if you'd rather click.
It lives in a closed shadow root, so no page can restyle it or find it, and only the panel itself takes clicks — never the space around it. It's kept out of screenshots automatically.
ChromeBoost is a Chrome extension plus an MCP server. The plugin ships both —
there's nothing to clone, no npm install, no build step. You need
Chrome and Claude Code.
In Claude Code, run these one at a time — not as a single paste.
Then restart Claude Code so the MCP server starts.
This copies the extension to your Downloads folder and opens it, so it's one click away in Chrome's folder picker.
chrome://extensions into Chrome's address bar
— chrome:// links can't be clicked from a page.
ChromeBoost-Extension from Downloads — or just drag the folder onto the page.Leave that folder where it is. Chrome reads it on every launch, so moving or deleting it breaks the extension. Chrome will also flag it as a developer-mode extension — expected, since ChromeBoost isn't on the Web Store and runs from files you can read.
Click the ChromeBoost icon in the toolbar. Your session shows up with a green dot — hit Use this window to bind the Chrome window you want driven. Then try:
It wants the repo on its own, without the /plugin prefix:
owner/repo shorthand clones over SSH. Use the HTTPS URL instead:
Or set CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1 to make shorthand use HTTPS.
The extension is loaded but hasn't connected. Hit the reload arrow on the ChromeBoost
card in chrome://extensions — it reconnects on a backoff timer and
a reload forces it immediately.
Run /chromeboost-setup again — it overwrites the Downloads folder
in place, so you only need to hit reload in chrome://extensions.
Clone github.com/lordamdal/chromeboost,
run npm install && npm run build, and load
packages/extension/dist. The local WebSocket bridge takes the first
free port in 7970–7980, so several sessions run side by side.
You won't type any of these — Claude picks them. Listed so you know exactly what it's able to do on your machine.
| Tool | What it's for |
|---|---|
hover | Park the cursor so hover-only menus, row actions and tooltips appear |
drag | Press, move, release — sliders, canvas, map panning, drag-and-drop reordering |
probe_point | Report what's actually painted at a spot; diagnoses a click that went nowhere |
hud | Move, dock, collapse or hide the on-page panel |
click_element · click_at_coordinates | Trusted clicks that hit-test first, scroll clear of fixed bars, and get past overlays |
fill_input · fill_form · type_text | Real keystrokes through React, contenteditable, CodeMirror, Monaco, TipTap |
get_page_text · find_text · get_form_fields | Read the page, including closed shadow DOM; secrets redacted by default |
open_page · wait_for · list_tabs · switch_to_tab | Navigate, wait for things to appear, work across tabs |
take_screenshot · download_file · read_attachment | Capture steps, pull authenticated files, read PDFs and docs |
fetch_url · inspect_request_headers | Requests with your real cookies, bypassing page CSP |
highlight_region · wait_for_click | Hand control back to you — 2FA, passwords, payments, anything that publishes |
write_to_env | Put a captured key straight into .env |