Riffle MCP server

Riffle

Your agent uses the web as text, not screenshots.

Pages as text. Only what changed comes back. Several steps per call. Fully interactive, without a mouse or keyboard.

Why Riffle · Install · Use · Stay signed in · Safe by default · Releases

npx @noetive/riffle init

Claude Code · Cursor · GitHub Copilot · Kiro · Antigravity · any MCP client

Riffle is an MCP server and CLI. It runs Chrome with the page's JavaScript and hands your agent each page as text it can read and act on.

Why Riffle

Your agent sees what a person sees

What is covered, hidden, primary, disabled, struck through or truncated is written on the line, in words. No HTML, no CSS, no images: Riffle turns the page into short lines of text, so your agent reads these facts instead of guessing them from markup or pixels.

modal d1 "Cookie preferences" covers=page
  text "We use cookies."
  button b1 "Accept all" primary
  button b2 "Reject"
main covered-by=d1
  h1 "Your cart"
  list
    item "Trail shoe 42" "€89" strike "€69" red | "Qty" | spinbutton f1 "Qty" ="1" | button b3 "Remove"
  button b4 "Checkout"

The cart sits behind a cookie dialog, the old price is struck through and the new one is red. covered-by=d1 tells your agent the checkout button sits behind the dialog.

Context that lasts the whole task

After the first view, every reply carries only what changed. Click "Reject" on the page above and this is the whole reply:

- d1
- text "We use cookies."
- b1
- b2
~ main -covered-by=d1

Every view also takes a token budget, so a long page costs what you set: view interactive budget=800.

Fewer turns, less waiting

Your agent writes a program, one step per line, and Riffle runs all of it in one call. A whole sign-in is one turn:

goto https://shop.example/login
try click "Reject"
fill "Email" "ralph@example.com"
fill "Password" $secret:shop
click "Sign in"
expect url ~ /account
view interactive budget=800

After each step Riffle waits until the page has answered and gone quiet instead of sleeping for a fixed time: about a second after a click, a few seconds after a page loads. The program stops at the first step that fails and says which step and why.

No mouse, no keyboard, no screenshots

Your agent names what to act on by the words on it or by a ref from a view: click "Reject", fill "Email" "ralph@example.com", click b4. Your agent never steers a pointer, aims at coordinates or reads an image. fill sets a whole field in one step, and press Enter names the key. When a step can't happen, the reply says what is in the way:

failed line 1: click "Checkout": blocked: b4 covered-by d1 "Cookie preferences"

Your agent's next program closes the dialog first.

Instead of

Shortened. The full README is on GitHub.