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

Related MCP server: Aperture

Install

npx @noetive/riffle init --client claude-code

Other editors: --client cursor, copilot, kiro, antigravity. Run init with no --client and it configures the editor it finds. --dry-run shows the change without writing it. --keep-state NAME keeps the agent signed in between runs; see Stay signed in.

Any MCP client can use this entry:

{
  "mcpServers": {
    "riffle": { "command": "npx", "args": ["-y", "@noetive/riffle", "mcp"] }
  }
}

On native Windows, run npx through cmd: "command": "cmd", "args": ["/c", "npx", "-y", "@noetive/riffle", "mcp"]. init writes this form for you.

Or install the binary: go install github.com/noetive/riffle/cmd/riffle@latest, or take one from the releases.

Riffle needs Chrome or Chromium on the machine. Check with:

riffle doctor

Riffle runs its own throwaway Chrome. It never touches your saved passwords or keychain, and it cleans up after itself when it stops.

Use

Your agent gets two tools. browser_run takes a program, one step per line, like the sign-in under Fewer turns, less waiting. browser_view reads the page.

Not sure of the syntax? Ask for view help, or run riffle grammar.

Steps: goto back forward click dblclick fill select check press hover scroll upload wait expect dialog try view eval. Views: outline, interactive, read, table REF, find "text", expand REF, net, unseen.

From a shell:

riffle run  < program.txt      # run a program
riffle view interactive        # read the current page
riffle close                   # end the session and its browser
riffle stop                    # end Riffle and every session in it

Each editor's agent has a browser of its own, so two windows never drive the same page. Its MCP server names its session on stderr at start; riffle view -s NAME shows that agent's page. Entries that pass the same -s NAME, as in riffle mcp -s shared, share one browser.

Riffle keeps browsers from piling up on your machine. At most four run at once, however many agents use Riffle (riffle serve -max-browsers N to change it). An agent that needs another waits for one to end, then is told which sessions hold them. A session unused for 15 minutes ends, and its next use starts on a blank page and says so. An editor's session ends when the editor stops its MCP server, and a Riffle with nothing to do exits after 30 minutes.

Stay signed in

Name a session in your editor's entry and Riffle keeps its sign-ins. The agent signs in once, and the next time the editor starts it is still signed in.

npx @noetive/riffle init --client claude-code --keep-state shop

or, by hand:

{ "mcpServers": { "riffle": { "command": "npx", "args": ["-y", "@noetive/riffle", "mcp", "-keep-state", "shop"] } } }

• What survives. Cookies and site storage are saved after every call, so they outlast the editor restarting, Riffle stopping and the browser crashing. Starting again sends nothing to any site: the first request is the agent's own.

• Where it is kept. At start Riffle names the file in its MCP server log. Only you can read it, but whoever can read it can use those sign-ins: keep it out of shared backups and synced folders. To sign the session out of everything, close the editors that use the name and delete the file. A Riffle upgraded from an earlier release may also have the earlier copy, in a riffle.state folder beside it: delete that one too.

• One name, one browser. Editors whose entries use the same name share one browser and one set of sign-ins. Give each project its own name to keep them apart. A name is lower case letters, digits, - and _.

• Limits still hold. A session started with -allow-origin gets back only the sign-ins of the sites it allows, and none for private addresses unless -allow-private is set. What a limit leaves out stays kept for a session allowed it.

• What is not kept. Storage is kept for the site the agent is on when each call ends, not for a site it only passed through within a call, nor for a frame from another site. Storage is kept for up to 50 sites; the one used longest ago goes first. A site whose storage cannot be put back is reported to the agent as event unrestored, and the session starts without it; it stays kept.

What you can do with it

• Pull a table out of a page. table REF returns rows as TSV, whatever the markup.

• See what the page calls. view net lists the fetch and XHR requests behind it.

• Test your own web app. Drive localhost the way a user would, with real page JavaScript running. Start Riffle with -allow-private so it may reach local addresses: riffle mcp -allow-private in your editor's entry, or riffle serve -allow-private from a shell, which takes the place of a Riffle already running without it, ending its sessions.

• Stay signed in between runs. Sign in once with -keep-state NAME, and the agent is still signed in the next time the editor starts. See Stay signed in.

Pages run in real time, as in any browser, and keep running while your agent thinks: a countdown, an expiring sign-in or a panel that opens later moves on between calls, and the next reply tells what changed.

Riffle also says so when a step did not do what it asked, such as a form that was refused, so the agent can fix it and go on.

Safe by default

• Secrets are written as $secret:name, tied to one origin, and never shown back to the agent.

• Text a person couldn't see on the page is counted, and shown only when you ask, inside a marker that says it's page data.

• A control whose hidden label contradicts the words on it is flagged name-differs, which also gives the words a person sees.

• Page text is data. Replies quote it, so a page can't pass it off as your instructions.

• Local and private network addresses are off until you turn them on with -allow-private, so a page cannot send the agent into your network or a cloud metadata service.

• Uploads and eval are off until you turn them on. -allow-origin and -upload-dir limit where a session can go and what it can send. A Riffle started without -allow-private, -allow-origin, -allow-eval or -upload-dir, from a shell or after a crash, keeps those it last ran with. Naming any of them sets all four, so -allow-private=false alone goes back to the defaults.

• Sign-ins end with their session, unless you keep them with -keep-state NAME. Kept sign-ins are in a file only you can read; delete it to forget them. A session gets back only the sign-ins its limits allow.

What it isn't

• Not a screenshot tool. If the answer is only in the pixels, Riffle can't give it to you.

• Not a way around bot protection or CAPTCHAs. It doesn't try. Every request says it comes from Riffle: the user agent ends in Riffle/<version>.

• Not a test framework for visual regressions. Use a tool built for that.

• Chrome and Chromium only. Content inside iframes is reported as content-not-shown, not read.

• Links that open a mail or phone app do nothing, and a new window opens in the same tab, so flows that need a pop-up do not work.

Works with the rest of Noetive

Riffle runs alone. It also pairs with another Noetive service:

• Semantik is a semantic message broker. An agent that reads a page can publish what it found, and its peers find it by meaning, with no topic names to agree on.

Develop

make build
make test
make lint
go test -tags integration ./...   # drives a real Chrome

See CONTRIBUTING.md. Report security issues as described in SECURITY.md.

License

ISC. Copyright (c) 2026 Noetive.io