Wink

Wink versus Playwright

The same agent ran the same eleven tasks in both browsers. Wink was faster and used fewer tokens on every one, and never needed more steps. On six real websites, with two Claude models, it needed 46% fewer steps, 16% less time and 46% fewer tokens.

21% fewer tool calls

28% less time

46% fewer tokens

33/33 runs done (Playwright 33/33)

Seconds per task

Wink Playwright MCP

Median seconds per task, Wink then Playwright: Sign-up form 11.5 and 13.4; To-do app 9.3 and 13.6; Wikipedia 8.8 and 12.7; Page dialogs 9.4 and 12.8; File upload 7.2 and 13.0; Drag and drop 6.9 and 9.6; Mock an API 9.2 and 10.6; Go offline 9.1 and 15.3; Cookies and storage 7.3 and 11.6; Save a PDF 6.9 and 8.1; Record a video 9.1 and 11.5. 0 5 10 15 20 s Sign-up form 11.5 13.4 To-do app 9.3 13.6 Wikipedia 8.8 12.7 Page dialogs 9.4 12.8 File upload 7.2 13.0 Drag and drop 6.9 9.6 Mock an API 9.2 10.6 Go offline 9.1 15.3 Cookies and storage 7.3 11.6 Save a PDF 6.9 8.1 Record a video 9.1 11.5 The same seconds per task, for narrow screens. Sign-up form 11.5 13.4 To-do app 9.3 13.6 Wikipedia 8.8 12.7 Page dialogs 9.4 12.8 File upload 7.2 13.0 Drag and drop 6.9 9.6 Mock an API 9.2 10.6 Go offline 9.1 15.3 Cookies and storage 7.3 11.6 Save a PDF 6.9 8.1 Record a video 9.1 11.5

Every task, side by side

Average tool calls, seconds and input tokens per task, Wink and Playwright MCP
TaskTool callsSecondsTokens in
WinkPlaywrightWinkPlaywrightWinkPlaywright
Sign-up formFill it in, submit, report the error and the failing request6711.513.451k71k
To-do appAdd three, complete one, delete one, filter4129.313.651k93k
WikipediaSearch, open the article, read a date from the infobox348.812.750k87k
Page dialogsAnswer a confirm and a prompt779.412.863k139k
File uploadAttach two files457.213.037k104k
Drag and dropA card between columns, a knob along a rail556.99.637k70k
Mock an APIMake a broken API answer, then sign up569.210.651k53k
Go offlineOffline and back, reading the page each time579.115.349k121k
Cookies and storageReport what a page stored557.311.637k68k
Save a PDFPrint the page to a file226.98.136k50k
Record a videoFilm two clicks669.111.549k86k
All 11526695132510k939k

On real websites

Six public sites, with two Claude models, both browsers headless and private. Medians of two runs. Wink was ahead on steps, time and tokens with both models: with Sonnet 16%, 5% and 30% ahead, with Haiku 56%, 26% and 53%. A few sites came within a few tenths of a second either way.

Sonnet 5.5

Real websites with Sonnet 5.5: runs done, median tool calls, seconds and input tokens
SiteDoneTool callsSecondsTokens in
WinkPlaywrightWinkPlaywrightWinkPlaywrightWinkPlaywright
Sauce Demo checkoutLog in, fill a cart, check out, read the total2/22/211.58.38.425k42k
Dynamic loadingStart a slow load, read what appears2/22/21311.413.324k68k
TodoMVC (React)Add three, complete one, clear completed2/22/2126.17.825k50k
Hacker NewsThe top story and its points2/22/23.528.98.655k50k
GitHubA repository's stars2/22/2227.56.936k34k
MDNFind one fact in long documentation2/22/2227.87.737k43k
All12/1212/1210.512.55053203k288k

Haiku 5.5

Real websites with Haiku 5.5: runs done, median tool calls, seconds and input tokens
SiteDoneTool callsSecondsTokens in
WinkPlaywrightWinkPlaywrightWinkPlaywrightWinkPlaywright
Sauce Demo checkoutLog in, fill a cart, check out, read the total2/22/22117.413.841k154k
Dynamic loadingStart a slow load, read what appears2/22/22.54.510.812.439k64k
TodoMVC (React)Add three, complete one, clear completed2/22/24127.614.161k209k
Hacker NewsThe top story and its points2/22/233.56.05.852k71k
GitHubA repository's stars2/22/2225.24.739k35k
MDNFind one fact in long documentation2/22/22.53.56.27.246k65k
All12/1212/121636.54358278k598k

Many at once

Private headless sessions running side by side on one Mac (an M3 with 16 GB), each adding ten to-dos through its MCP server, scripted, with no model in the loop. Every session in both browsers finished.

Time for all sessions to finish and peak memory, Wink and Playwright MCP
Sessions at onceAll finished inPeak memory
WinkPlaywrightWinkPlaywright
12.0 s6.6 s0.6 GB1.3 GB
42.2 s8.1 s2.7 GB4.1 GB
83.3 s9.5 s5.2 GB7.5 GB

What each can do

Capabilities of Wink and Playwright MCP
CapabilityWinkPlaywright MCP
Click, type, keys, forms, hover, scrollYesYes
Drag and drop, file upload, page dialogsYes; uploads ask you firstYes
Console, network requests in fullYesYes
Mock the network, go offlineYesYes, with an option turned on
Cookies and storageYes; values hidden outside localhostYes, with an option turned on
Phone sizes, dark mode, PDF, videoYesYes, partly with options turned on
Check the page, write a testYes; checks verified before they're writtenYes
Step-by-step traceYes: a timeline of steps with screenshots, console and networkYes, with an option turned on
Headless, isolated, several at onceYes: a private throwaway Wink per agent, on a Mac; 8 at once about 3× faster, with 31% less memoryYes, on any platform
Runs in CIYes, on macOS runners; its tests on anyYes, anywhere
Firefox and WebKit enginesThrough the tests it writes (checked on every change, on Linux)Yes
A whole known flow in one callYes: batch, by selector or label, every step guardedYes: run_code, which Playwright calls unsafe
Your own browser, your loginsYes, with your consentWith a browser extension
You see it work, pause it, stop itYesNo
Asks before it sends, buys, deletes or uploadsYesNo
Set-up in Claude CodeNone: Wink adds itselfNode and a config line
Tool descriptions the agent rereads every turn25 KB (37 tools)42 KB (60 tools) with the same options

Why it takes fewer steps

  • It says what changed

    After each click Wink reports only the difference: a box now checked, a button gone, a counter now reading “1 item left”. The agent rarely needs to look at the whole page again.

  • Its references hold

    When a page redraws a list, Wink's handles follow their elements. In our runs, Playwright's went stale once and the agent had to start that step over.

  • It reads what you ask

    Ask for one fact and Wink returns the rows that mention it, not the whole article. Fill a form in one call, or run a sequence in one batch.

  • It keeps the cache warm

    Wink describes the same abilities in 37 tools where Playwright needs 60, about 40% less text to reread every turn, and identically each time, so the agent's prompt cache keeps hitting.

What only Wink does

  • Your browser, in plain sight

    The agent works in a browser you can see, with a glow, a cursor and a log. Click into its page and it pauses. ⇧⌘. stops every agent.

  • It asks first

    Wink asks before an agent sends, buys, deletes, connects or uploads anything. Your own app on localhost is exempt, because you're building it.

  • Nothing to set up

    Install Wink and Claude Code can use it, with no Node and no config. Wink opens only when an agent first browses, and keeps working if your screen locks.

  • It writes the test

    Ask the agent to export what it did, and Wink writes a Playwright test with checks verified on the live page. The tests from our runs pass in Chromium, Firefox and WebKit.

When Playwright is still the tool

Wink runs on Macs, on Chromium. For test suites on Linux or Windows servers, in Firefox's and Safari's own engines, or headless by the hundred on a server farm, use Playwright. Wink writes the Playwright tests those servers run, so you can use both.

How we measured

Claude Code 2.1.322 with Claude Sonnet 5.5, run headless (claude -p) with one browser connected and its other tools off. Wink 0.8.1 in its headless mode (a private, throwaway instance per session) against Playwright MCP 0.0.85 (headless, isolated, Chrome), with Playwright's network, storage, PDF and DevTools options turned on so it had a tool for every task. The eleven tasks ran with Claude Sonnet 5.5; the six real websites with Sonnet 5.5 and Claude Haiku 5.5. Each of the eleven tasks ran three times per browser and each website twice per browser and model, taking turns, on 11 November 2026, all on the build released as 0.8.1; figures are the median run, so one slow run can’t tip them. Every run completed its task. Playwright’s video task was run again after we found we had removed the ffmpeg download it needs for video; the rerun covered both browsers. Tokens are input tokens, cached ones included. Ten of the eleven tasks used small test pages on the same Mac and one the live Wikipedia; the websites were Sauce Demo, The Internet, TodoMVC, Hacker News, GitHub and MDN. Every prompt was the same for both. The tests Wink exported pass in Playwright on Chromium, Firefox and WebKit. Wink's code is open source: github.com/wink-browser/wink.

Give your agent a better browser.

Free, open source, and on your Mac in a wink.

Download for Mac

macOS 14 or later · Apple silicon