Let AI agents use any website.
Your AI agent can log in, fill forms, click buttons and read any page — just like a person would, but 11x faster than headless Chrome in our tests. No screenshots, no fragile code.
Start in one command
$ npx memora-browser --port 9222The npm package is coming soon. Until then, run it from source.
- Open source, Apache-2.0
- Playwright, Puppeteer, Selenium
- MCP server built in
Works with
- Playwright
- Puppeteer
- Selenium
- Claude Code
- Cursor
- Windsurf
Lighter than Chrome. Measured.
Same laptop, same 200 pages. Memora Browser doesn’t draw anything, so it loads each page faster and uses a fraction of the memory.
11x faster
Time to load 200 pages
11x less memory
Memory with 10 tabs open
200 local test pages, Windows 11 laptop, median of three runs, October 2026. Timed after each browser had started. Chrome is Playwright’s headless shell. Method and more results
How it works, in three steps.
Normal browsers draw pages for people. Memora Browser reads them for machines, and turns each page into a short list your agent can act on.
@1 Give it a URL
Start Memora Browser and point it at any page: a website, a dashboard or your own app on localhost.
@2 It labels the page
It runs the page’s JavaScript, finds every button, link and field, and gives each one a number.
@3 Your agent acts
The agent reads the short list and says what to do by number. Memora clicks or types on the right element.
Use it with the tools you already have.
Existing scripts only change how they connect. AI coding assistants get it as an MCP server. Or call it directly from Node.js.
# Start the automation server (CDP on port 9222)
npx memora-browser --port 9222
# Or turn one page into an action map
npx memora-browser fetch https://example.com --dump action_mapnpm package coming soon — until then, run from source (see Docs).
Everything an agent needs on the web.
Each feature has its own page with examples. Pick one to dig in.
Action map
Every link, button and field on a page gets a number like @1 or @2, so an agent can act on it without writing selectors.
Page forking
Copy a page with its DOM, cookies and storage into a new tab, then try a different path without loading it again.
MCP server
Give Claude Code, Cursor and other MCP clients a browser with 12 tools: browse, act, extract, fork and more.
CDP & WebDriver
Connect Playwright and Puppeteer over CDP, or Selenium over WebDriver. Existing scripts only change how they connect.
Stealth profile
A consistent Chrome 130 profile: matching headers, client hints and navigator values, with navigator.webdriver set to false.
Fetch & dump
One command turns a URL into Markdown, HTML, structured data, a semantic tree or an action map.
Why not just use Chrome?
Chrome is built for people. When an agent drives it, most of the work goes into things it never looks at.
| Topic | Headless Chrome | Memora Browser |
|---|---|---|
| Built for | People looking at a screen | AI agents and scripts |
| Draws the page | Yes — layout, fonts, painting | No — DOM and JavaScript only |
| What the agent reads | Raw HTML or screenshots | Numbered actions and clean Markdown |
| How the agent clicks | CSS or XPath it has to write | By number: click @3 |
| Screenshots and PDFs | Yes | No — use Chrome when you need pixels |
Common questions.
What is Memora Browser?
How is it different from headless Chrome?
Does it work with Playwright, Puppeteer and Selenium?
Is it open source?
How do I install it?
Give your agent a browser it understands.
Start the server, connect your agent and let it use the web by number.