Published 4 min read

Testing agent-browser, Playwright and Chrome DevTools with Claude Code and Codex

I hand a lot of browser jobs to coding agents, and I wanted to find out which tool is fastest and cheapest for those jobs. So I had Claude Code and Codex do the same job in a signed-in web app with each tool, over its CLI with and without its skill and over its MCP server. The tools were mostly close on speed and cost. The big exception was agent-browser, which with Claude Code used 2.4 times as many input tokens over MCP as over its CLI with the skill.

Contents

I stuck to open-source tools, because I think building production workflows around proprietary ones, like the built-in browser in OpenAI’s desktop app or the Claude in Chrome extension, is a bad long-term strategy. Those tie you to one vendor’s agents and models, while all three tools I tested ran with both Claude Code and Codex, and work with any other agent harness. For a production workflow, it’s also important to be able to study a tool’s code, pin a version and fork it if it doesn’t fit your needs or its maintainers take it in a new direction.

Key results

  1. Speed had no clear winner, with one exception. Chrome DevTools over MCP with Claude Code was clearly faster than the other two tools.
  2. With Claude Code, MCP was the most expensive connection for every tool.
  3. agent-browser was cheaper over its CLI without its skill, with both Claude Code and Codex.
  4. A Python script did the same job faster than every agent run.
  5. Windows was the slowest place to run, slower than WSL on the same PC.

With Claude Code, every tool used the most input tokens over MCP

Claude Code (Opus 5.5, medium). The mark is each setup’s median of 12 runs, and each tick on its line is one run. Lower is better.

Show the Claude Code values
With Claude Code, every tool used the most input tokens over MCP
SetupInput tokens per run, medianlowesthighestTime per run, medianlowesthighest
agent-browser, CLI + skill652 thousand472 thousand1.09 million104 seconds82 seconds142 seconds
Playwright, CLI + skill332 thousand257 thousand400 thousand90 seconds74 seconds128 seconds
Chrome DevTools, CLI + skill341 thousand283 thousand508 thousand97 seconds79 seconds120 seconds
agent-browser, CLI, no skill304 thousand244 thousand395 thousand96 seconds73 seconds126 seconds
Playwright, CLI, no skill353 thousand251 thousand410 thousand99 seconds80 seconds125 seconds
Chrome DevTools, CLI, no skill348 thousand288 thousand401 thousand98 seconds81 seconds116 seconds
agent-browser, MCP1.59 million1.29 million1.99 million106 seconds87 seconds170 seconds
Playwright, MCP513 thousand417 thousand691 thousand98 seconds87 seconds119 seconds
Chrome DevTools, MCP517 thousand455 thousand592 thousand72 seconds62 seconds112 seconds
Note: Input tokens include cache reads. Time runs from launching the harness to its final message. Each measure and each harness has its own scale. Source: my study runs, 19 to 22 October 2026.

With Codex, agent-browser used the most input tokens over every connection

Codex (GPT 6 Astra, medium). The mark is each setup’s median of 12 runs, and each tick on its line is one run. Lower is better.

Show the Codex values
With Codex, agent-browser used the most input tokens over every connection
SetupInput tokens per run, medianlowesthighestTime per run, medianlowesthighest
agent-browser, CLI + skill426 thousand278 thousand505 thousand91 seconds70 seconds127 seconds
Playwright, CLI + skill318 thousand211 thousand399 thousand102 seconds70 seconds151 seconds
Chrome DevTools, CLI + skill197 thousand173 thousand257 thousand75 seconds68 seconds168 seconds
agent-browser, CLI, no skill289 thousand228 thousand335 thousand78 seconds59 seconds107 seconds
Playwright, CLI, no skill243 thousand209 thousand368 thousand91 seconds74 seconds160 seconds
Chrome DevTools, CLI, no skill233 thousand207 thousand283 thousand86 seconds69 seconds277 seconds
agent-browser, MCP414 thousand333 thousand473 thousand72 seconds59 seconds86 seconds
Playwright, MCP261 thousand188 thousand335 thousand88 seconds66 seconds117 seconds
Chrome DevTools, MCP238 thousand166 thousand365 thousand81 seconds60 seconds112 seconds
Note: Pooled across the three OSes, run times overlapped in every Codex matchup, so no tool was clearly faster. Each measure and each harness has its own scale. Source: my study runs, 19 to 22 October 2026.

The job

Two harnesses, three connections and three tools make 18 setups

Every run did the same steps in a signed-in web app

  1. Pick the right option from a menu.

  2. Open a report in a new tab and check it.

  3. Read its hundreds of rows and save them.

  4. Close the report’s tab.

  5. Go back to the start page.

Speed

Only Chrome DevTools over MCP with Claude Code was clearly faster than the other two tools

Every run’s time with Claude Code (Opus 5.5, medium) and Codex (GPT 6 Astra, medium), 12 runs per tool in each matchup. The other 5 matchups are in grey.

Strip chart of every run’s time, one dot per run, with a panel for each of the 6 matchups. Only Chrome DevTools over MCP with Claude Code was clearly faster than the other two tools. The time axis starts at 40 seconds, not at zero. Each run’s values are in the table of every run at the end of this post.

Skip the chart of every run

Claude Code, MCP

agent-browser 87 to 170 seconds, Playwright 87 to 119 seconds, Chrome DevTools 62 to 112 seconds. Chrome DevTools: 10 of its 12 runs were faster than every run of the other two tools in this matchup, left of the dashed line.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI + skill

agent-browser 82 to 142 seconds, Playwright 74 to 128 seconds, Chrome DevTools 79 to 120 seconds.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI, no skill

agent-browser 73 to 126 seconds, Playwright 80 to 125 seconds, Chrome DevTools 81 to 116 seconds.

agent-browser
Playwright
Chrome DevTools

Codex, CLI + skill

agent-browser 70 to 127 seconds, Playwright 70 to 151 seconds, Chrome DevTools 68 to 168 seconds.

agent-browser
Playwright
Chrome DevTools

Codex, CLI, no skill

agent-browser 59 to 107 seconds, Playwright 74 to 160 seconds, Chrome DevTools 69 to 277 seconds.

agent-browser
Playwright
Chrome DevTools

Codex, MCP

agent-browser 59 to 86 seconds, Playwright 66 to 117 seconds, Chrome DevTools 60 to 112 seconds.

agent-browser
Playwright
Chrome DevTools
Note: All panels share one time axis, which starts at 40s. A run past 200s sits on its right edge and shows its time. Source: my study runs, 19 to 22 October 2026.

Most of each run was waiting on the model

Median time per run in seconds, 12 runs per setup. Both harnesses use one scale.

Most of each run was waiting on the model
SetupMedian time per runPart inside the browser tool
Claude Code, agent-browser, CLI + skill104 seconds11 seconds
Claude Code, agent-browser, CLI, no skill96 seconds14 seconds
Claude Code, agent-browser, MCP106 seconds4 seconds
Claude Code, Playwright, CLI + skill90 seconds18 seconds
Claude Code, Playwright, CLI, no skill99 seconds16 seconds
Claude Code, Playwright, MCP98 seconds9 seconds
Claude Code, Chrome DevTools, CLI + skill97 seconds14 seconds
Claude Code, Chrome DevTools, CLI, no skill98 seconds17 seconds
Claude Code, Chrome DevTools, MCP72 seconds4 seconds
Codex, agent-browser, CLI + skill91 seconds3 seconds
Codex, agent-browser, CLI, no skill78 seconds3 seconds
Codex, agent-browser, MCP72 seconds3 seconds
Codex, Playwright, CLI + skill102 seconds12 seconds
Codex, Playwright, CLI, no skill91 seconds12 seconds
Codex, Playwright, MCP88 seconds13 seconds
Codex, Chrome DevTools, CLI + skill75 seconds9 seconds
Codex, Chrome DevTools, CLI, no skill86 seconds10 seconds
Codex, Chrome DevTools, MCP81 seconds3 seconds
Note: The browser tool took 3 to 18s of each setup’s median run. Source: my study runs, 19 to 22 October 2026.

Since every run passed its checks, I’d choose a tool by its token use and speed.

Split by OS

The charts above combine the three operating systems. Split by OS, each setup has 4 runs. I compared each pair of tools within each harness, connection and OS. In 15 of those 54 comparisons, every run of one tool was faster than every run of the other, and 9 of the 15 were on Windows.

With Claude Code, only Chrome DevTools over MCP on macOS was faster than both other tools

Claude Code (Opus 5.5, medium). Each dot is one run, 4 per tool in each panel, and further left is faster. A panel is in colour when every run of one tool was faster than every run of both others, and a dashed line marks the fastest run of the other two. A note under a panel names any tool whose every run was faster than every run of another.

Strip chart of Claude Code’s run times, one dot per run, with a panel per connection and OS. With Claude Code, only Chrome DevTools over MCP on macOS was faster than both other tools. The time axis starts at 40 seconds, not at zero. Each run’s values are in the table of every run at the end of this post.

Skip the chart of Claude Code’s runs on each OS

Claude Code, CLI + skill, macOS

agent-browser 97 to 142 seconds, Playwright 78 to 128 seconds, Chrome DevTools 79 to 112 seconds.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI + skill, WSL

agent-browser 82 to 96 seconds, Playwright 74 to 102 seconds, Chrome DevTools 79 to 119 seconds.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI + skill, Windows

agent-browser 93 to 129 seconds, Playwright 89 to 100 seconds, Chrome DevTools 100 to 120 seconds. Playwright was faster than Chrome DevTools.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI, no skill, macOS

agent-browser 89 to 103 seconds, Playwright 80 to 102 seconds, Chrome DevTools 84 to 98 seconds.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI, no skill, WSL

agent-browser 73 to 92 seconds, Playwright 94 to 116 seconds, Chrome DevTools 81 to 110 seconds. agent-browser was faster than Playwright.

agent-browser
Playwright
Chrome DevTools

Claude Code, CLI, no skill, Windows

agent-browser 83 to 126 seconds, Playwright 101 to 125 seconds, Chrome DevTools 102 to 116 seconds.

agent-browser
Playwright
Chrome DevTools

Claude Code, MCP, macOS

agent-browser 97 to 107 seconds, Playwright 91 to 119 seconds, Chrome DevTools 64 to 85 seconds. Chrome DevTools was faster than agent-browser and Playwright.

agent-browser
Playwright
Chrome DevTools

Claude Code, MCP, WSL

agent-browser 87 to 131 seconds, Playwright 87 to 116 seconds, Chrome DevTools 62 to 107 seconds.

agent-browser
Playwright
Chrome DevTools

Claude Code, MCP, Windows

agent-browser 133 to 170 seconds, Playwright 91 to 105 seconds, Chrome DevTools 69 to 112 seconds. Playwright and Chrome DevTools were faster than agent-browser.

agent-browser
Playwright
Chrome DevTools
Note: All panels share one time axis, which starts at 40s. Source: my study runs, 19 to 22 October 2026.

With Codex on Windows, agent-browser was faster than both other tools over every connection

Codex (GPT 6 Astra, medium). Each dot is one run, 4 per tool in each panel, and further left is faster. A panel is in colour when every run of one tool was faster than every run of both others, and a dashed line marks the fastest run of the other two. A note under a panel names any tool whose every run was faster than every run of another.

Strip chart of Codex’s run times, one dot per run, with a panel per connection and OS. With Codex on Windows, agent-browser was faster than both other tools over every connection. The time axis starts at 40 seconds, not at zero. Each run’s values are in the table of every run at the end of this post.

Skip the chart of Codex’s runs on each OS

Codex, CLI + skill, macOS

agent-browser 84 to 95 seconds, Playwright 70 to 106 seconds, Chrome DevTools 68 to 73 seconds. Chrome DevTools was faster than agent-browser.

agent-browser
Playwright
Chrome DevTools

Codex, CLI + skill, WSL

agent-browser 70 to 127 seconds, Playwright 78 to 104 seconds, Chrome DevTools 72 to 78 seconds.

agent-browser
Playwright
Chrome DevTools

Codex, CLI + skill, Windows

agent-browser 91 to 106 seconds, Playwright 127 to 151 seconds, Chrome DevTools 140 to 168 seconds. agent-browser was faster than Playwright and Chrome DevTools.

agent-browser
Playwright
Chrome DevTools

Codex, CLI, no skill, macOS

agent-browser 68 to 91 seconds, Playwright 74 to 106 seconds, Chrome DevTools 71 to 87 seconds.

agent-browser
Playwright
Chrome DevTools

Codex, CLI, no skill, WSL

agent-browser 59 to 78 seconds, Playwright 79 to 82 seconds, Chrome DevTools 69 to 107 seconds. agent-browser was faster than Playwright.

agent-browser
Playwright
Chrome DevTools

Codex, CLI, no skill, Windows

agent-browser 87 to 107 seconds, Playwright 125 to 160 seconds, Chrome DevTools 138 to 277 seconds. agent-browser was faster than Playwright and Chrome DevTools.

agent-browser
Playwright
Chrome DevTools

Codex, MCP, macOS

agent-browser 67 to 86 seconds, Playwright 66 to 117 seconds, Chrome DevTools 60 to 89 seconds.

agent-browser
Playwright
Chrome DevTools

Codex, MCP, WSL

agent-browser 61 to 68 seconds, Playwright 70 to 89 seconds, Chrome DevTools 63 to 81 seconds. agent-browser was faster than Playwright.

agent-browser
Playwright
Chrome DevTools

Codex, MCP, Windows

agent-browser 59 to 82 seconds, Playwright 87 to 111 seconds, Chrome DevTools 84 to 112 seconds. agent-browser was faster than Playwright and Chrome DevTools.

agent-browser
Playwright
Chrome DevTools
Note: All panels share one time axis, which starts at 40s. A run past 200s sits on its right edge and shows its time. Source: my study runs, 19 to 22 October 2026.

With Codex, agent-browser still used the most input tokens over its CLI with the skill on every OS. These results come from 4 runs per setup on each OS, on one Mac and one PC, so a rerun or another machine might not match them.

Token use over MCP

agent-browser over MCP with Claude Code was the most expensive of all 18 setups. It made about as many model calls as over its CLI with the skill, but each call used more input tokens.

Each tool’s model calls were largest over MCP with Claude Code

Median input tokens per model call, 12 runs per setup. Both harnesses use one scale.

Each tool’s model calls were largest over MCP with Claude Code
SetupMedian input tokens per model call
Claude Code, agent-browser, CLI + skill30.1 thousand
Claude Code, agent-browser, CLI, no skill15.9 thousand
Claude Code, agent-browser, MCP72.7 thousand
Claude Code, Playwright, CLI + skill20.7 thousand
Claude Code, Playwright, CLI, no skill17.6 thousand
Claude Code, Playwright, MCP24.5 thousand
Claude Code, Chrome DevTools, CLI + skill19.5 thousand
Claude Code, Chrome DevTools, CLI, no skill18.6 thousand
Claude Code, Chrome DevTools, MCP28.8 thousand
Codex, agent-browser, CLI + skill30.4 thousand
Codex, agent-browser, CLI, no skill19.9 thousand
Codex, agent-browser, MCP31.9 thousand
Codex, Playwright, CLI + skill20.6 thousand
Codex, Playwright, CLI, no skill17.2 thousand
Codex, Playwright, MCP18.9 thousand
Codex, Chrome DevTools, CLI + skill18.3 thousand
Codex, Chrome DevTools, CLI, no skill17.9 thousand
Codex, Chrome DevTools, MCP19.6 thousand
Source: my study runs, 19 to 22 October 2026.

With Codex, no connection was consistently cheaper.

agent-browser’s skill

agent-browser’s skill is a stub: it tells the agent to run agent-browser skills get core, which prints a long guide that stays in context for the rest of the run.

agent-browser’s skill puts the most in the agent’s context

Each square is 1,000 characters.

agent-browser’s skill puts the most in the agent’s context
ToolCharacters in its skillCharacters in the guide its skill loads
agent-browser3,45537,980
Playwright15,422None
Chrome DevTools12,031None
Source: the SKILL.md in each tool’s official skill at the version I tested, and the guide that agent-browser’s skill loads, 19 to 22 October 2026.

Over its CLI, dropping the skill cut agent-browser’s input tokens by 53% with Claude Code and 32% with Codex, while the other two tools’ CLI runs showed no consistent gap. In 67 of the 72 CLI runs without a skill, the agent read the tool’s --help instead. For a job like this one, I’d try agent-browser without its skill.

A script instead of an agent

Between agent runs, a Python script did the same job over CDP, using no model calls and no tokens at run time.

The script’s slowest run was faster than the fastest agent run

Time of every run in seconds: 258 runs of the script, and 108 agent runs with each harness.

The script’s slowest run was faster than the fastest agent run
What ranRunsFastestMedianSlowest
Python script over CDP2581.4 seconds1.5 seconds10.3 seconds
Claude Code (Opus 5.5, medium)10862 seconds97 seconds170 seconds
Codex (GPT 6 Astra, medium)10859 seconds84 seconds277 seconds
Note: The script’s times don’t include writing it. Source: my study runs, 19 to 22 October 2026.

An earlier Claude Code session did this job with agent-browser and left notes, and the session I built the study with turned those notes into the script. Then I broke the script on purpose, and an agent running it finished the job itself and proposed three fixes, which went into the script.

For a job you repeat, I’d let an agent write a script and call the agent back when the script breaks

  1. An agent does the job once and keeps notes.

  2. An agent turns the notes into a script.

  3. The script runs the job on its own.

  4. When the script breaks, an agent finishes the job and fixes the script.

Windows and WSL

Windows runs spent the most time inside the browser tool

Median time per run in seconds, 36 runs per harness on each OS. Both harnesses use one scale.

Windows runs spent the most time inside the browser tool
Harness and OSMedian time per runPart inside the browser tool
Claude Code, macOS97 seconds13 seconds
Claude Code, WSL92 seconds10 seconds
Claude Code, Windows107 seconds15 seconds
Codex, macOS79 seconds8 seconds
Codex, WSL77 seconds7 seconds
Codex, Windows109 seconds24 seconds
Note: WSL and Windows ran on the same PC. Source: my study runs, 19 to 22 October 2026.

Each tool’s CLI commands took longest on Windows

Median seconds per CLI command, over both harnesses’ CLI runs on each OS.

Each tool’s CLI commands took longest on Windows
Tool and OSMedian time per CLI command
agent-browser, macOS0.2 seconds
agent-browser, WSL0.1 seconds
agent-browser, Windows0.5 seconds
Playwright, macOS0.7 seconds
Playwright, WSL0.7 seconds
Playwright, Windows1.4 seconds
Chrome DevTools, macOS0.5 seconds
Chrome DevTools, WSL0.6 seconds
Chrome DevTools, Windows1.9 seconds
Source: my study runs, 19 to 22 October 2026.

On a Windows PC, I’d run agents in WSL.

Limits

I measured one job in one web app, at medium effort, with 12 runs per setup, so another job could order the tools differently. I report each harness on its own because they ran different models at different prices.

The Windows and WSL result comes from one PC and the versions listed below. Every number pools the three OSes unless I split it.

How I ran the study

Show the method

I used these versions:

  • Claude Code 2.1.303 with claude-opus-5-5, medium effort
  • Codex 0.163.0-alpha.11 with gpt-6-astra, medium effort
  • agent-browser 0.39.1 (Vercel Labs), the CLI and its built-in MCP server
  • Playwright CLI 0.1.22 and Playwright MCP 0.0.83 (Microsoft)
  • Chrome DevTools for agents 1.11.0 (Google), its chrome-devtools CLI and chrome-devtools-mcp server
  • Google Chrome 155, stable release, headless, with one study profile per operating system

The CLI + skill setups got each tool’s official skill folder, unchanged, from its GitHub repository at the version above: agent-browser, Playwright CLI and Chrome DevTools. Each tool’s documentation tells you to install that folder.

A run passed only if it saved rows that matched a reference exactly, left the app in the right state, used only its assigned browser tool, never asked me for input and finished within 30 minutes.

I ran the study from 19 to 22 October 2026. Every run was a fresh harness session with empty settings, no other skills or MCP servers and no project instruction files. Each got the same instructions, started with the app open and signed in, and read the report’s table with the same JavaScript. Each number is a setup’s median over its 12 runs unless I say otherwise. I took each measure’s median on its own, so a median run can mix values from different runs.

Token counts come from each harness’s own logs. Cache reads made up 93.9% of input tokens with Claude Code and 90.5% with Codex. The tables also show output tokens, which include reasoning.

Claude Code reports what each run would cost through the API, and I priced Codex’s tokens at list prices. Neither is what you’d pay on a subscription plan. Within one harness, cost roughly follows tokens, so I chart tokens and give costs in the text and the tables.

The web app’s name and the report’s contents stay private, so this post reports only derived measurements.

Every setup’s median run

Claude Code (Opus 5.5, medium)

12 runs per setup, 4 on each OS.

Every setup’s median run with Claude Code (Opus 5.5, medium)
ConnectionTimeInput tokensCostIn browser toolOutput tokensModel callsBrowser commands
agent-browser
CLI + skill104s652K$0.4811s6.8K2240
CLI, no skill96s304K$0.2814s5.9K2038
MCP106s1.59M$1.024s8.7K2127
Playwright
CLI + skill90s332K$0.3218s5.8K1622
CLI, no skill99s353K$0.3316s6.6K2025
MCP98s513K$0.419s7.3K2115
Chrome DevTools
CLI + skill97s341K$0.3214s6.1K1722
CLI, no skill98s348K$0.3217s5.9K1924
MCP72s517K$0.404s5.7K1815
Source: my study runs, 19 to 22 October 2026.

Codex (GPT 6 Astra, medium)

12 runs per setup, 4 on each OS.

Every setup’s median run with Codex (GPT 6 Astra, medium)
ConnectionTimeInput tokensCostIn browser toolOutput tokensModel callsBrowser commands
agent-browser
CLI + skill91s426K$0.913s2.2K1427
CLI, no skill78s289K$0.683s1.9K1423
MCP72s414K$0.893s1.8K1321
Playwright
CLI + skill102s318K$0.8012s2.2K1616
CLI, no skill91s243K$0.6612s2.2K1418
MCP88s261K$0.6213s2.1K1418
Chrome DevTools
CLI + skill75s197K$0.539s1.9K1117
CLI, no skill86s233K$0.5810s1.9K1318
MCP81s238K$0.603s2.0K1215
Source: my study runs, 19 to 22 October 2026.

Every run

One row per run, with time in seconds and cost in US dollars.

Show the 216 runs
Every run
OSRunTime (s)Input tokensCached inputOutput tokensCost (USD)Model callsBrowser commands
Claude Code, agent-browser, CLI + skill
macOSC1116.8623,265584,1109,156$0.502355
macOSC297.4569,199534,2815,713$0.421939
macOSC3142.4682,163646,5849,044$0.522339
macOSC4132.31,093,8811,048,68410,628$0.682952
WSLC195.7471,985437,9356,209$0.411736
WSLC282.2543,769509,8006,843$0.441841
WSLC383.9486,134452,0686,763$0.421724
WSLC495.8734,957691,7656,075$0.502028
WindowsC1129.2681,670646,3816,191$0.462243
WindowsC2109.9706,145668,7828,985$0.542252
WindowsC392.7594,483562,4716,643$0.432146
WindowsC4119.6962,950917,4089,827$0.642535
Claude Code, agent-browser, CLI, no skill
macOSC1101.6341,587321,5067,700$0.352042
macOSC288.9304,730288,5995,877$0.281944
macOSC399.9346,352329,6347,201$0.322241
macOSC4103.0340,204323,6555,568$0.292138
WSLC191.5269,273254,0685,683$0.271834
WSLC273.0244,496229,5185,356$0.261635
WSLC391.3360,465343,8987,731$0.332333
WSLC476.0294,146276,9516,077$0.291834
WindowsC1106.8395,009377,0817,175$0.342340
WindowsC283.0304,086290,2265,861$0.281940
WindowsC3126.1301,055287,8795,877$0.282037
WindowsC4121.8270,652256,5835,848$0.271937
Claude Code, agent-browser, MCP
macOSC1105.41,619,5311,495,6008,756$1.092127
macOSC2107.21,403,4131,277,6088,323$1.051927
macOSC3101.61,545,5041,438,8129,155$1.002128
macOSC497.01,603,4181,489,6938,623$1.042028
WSLC1130.91,617,8431,514,3708,517$0.992321
WSLC295.91,498,5811,394,8317,947$0.962126
WSLC386.71,289,8031,208,2088,102$0.812020
WSLC4101.51,577,9321,492,4599,168$0.912325
WindowsC1133.11,426,3081,357,5189,130$0.982030
WindowsC2168.21,904,7871,792,9418,486$1.092425
WindowsC3143.81,864,5861,798,4078,950$1.052626
WindowsC4169.61,992,2301,881,76710,233$1.132727
Claude Code, Playwright, CLI + skill
macOSC1127.5399,690375,6656,560$0.351828
macOSC277.5304,364278,4525,187$0.291523
macOSC390.9316,991297,2555,279$0.291622
macOSC478.8328,600307,2136,040$0.321618
WSLC184.5363,244339,7286,221$0.331723
WSLC2101.6381,186356,2665,429$0.331620
WSLC376.6314,550294,7465,618$0.301620
WSLC474.0256,717236,7004,776$0.271320
WindowsC199.5393,098365,6806,947$0.371622
WindowsC289.3282,898261,9324,714$0.281420
WindowsC394.3335,328313,1746,791$0.341622
WindowsC4100.0379,986357,8007,002$0.351822
Claude Code, Playwright, CLI, no skill
macOSC197.2410,283386,8026,966$0.362130
macOSC285.4251,297231,7936,492$0.301518
macOSC3102.1374,628353,6257,407$0.352125
macOSC479.9292,361273,6275,868$0.291721
WSLC193.7319,087299,5706,488$0.311822
WSLC295.2325,868308,6707,380$0.331922
WSLC3116.4365,122344,8746,989$0.342024
WSLC497.1341,101322,2375,763$0.302025
WindowsC1107.6387,138364,9697,251$0.352127
WindowsC2125.1316,374300,4515,270$0.281925
WindowsC3101.2373,167352,3756,693$0.342021
WindowsC4115.9384,501364,5106,187$0.322226
Claude Code, Playwright, MCP
macOSC1108.4521,899487,8257,236$0.412117
macOSC2112.4506,912472,4587,584$0.422017
macOSC390.9528,905495,5887,384$0.412213
macOSC4118.5655,019621,9227,202$0.432713
WSLC186.7518,440499,4507,166$0.412115
WSLC293.1417,396386,8046,597$0.361917
WSLC3115.5690,872665,0079,646$0.522719
WSLC493.3449,468416,0547,157$0.391815
WindowsC1102.9580,763546,8307,959$0.442315
WindowsC291.3463,217430,2937,551$0.401916
WindowsC393.3455,884423,2047,787$0.401915
WindowsC4104.6481,955449,2646,662$0.392015
Claude Code, Chrome DevTools, CLI + skill
macOSC178.9332,835307,0576,041$0.311721
macOSC279.1326,953306,6485,782$0.311723
macOSC378.9348,502326,4336,102$0.331721
macOSC4111.8291,502271,6295,977$0.301518
WSLC1118.9508,343482,9619,068$0.432330
WSLC294.2312,100292,9794,667$0.281619
WSLC378.7283,404264,4435,767$0.291520
WSLC485.0378,445357,4466,385$0.331922
WindowsC1116.2355,723335,2326,265$0.321822
WindowsC2119.8367,891348,4035,998$0.321921
WindowsC3100.3328,594308,5136,534$0.321722
WindowsC4117.8498,954457,6266,890$0.461825
Claude Code, Chrome DevTools, CLI, no skill
macOSC197.7345,847326,1965,925$0.311924
macOSC284.3387,549366,5606,166$0.341920
macOSC383.9304,904286,4396,221$0.301824
macOSC489.2325,593304,4685,299$0.301723
WSLC197.3342,741323,2535,392$0.291922
WSLC2110.1401,020380,6867,951$0.362224
WSLC384.9373,668350,7455,923$0.331824
WSLC480.9389,920367,8175,633$0.321924
WindowsC1102.3346,841326,6385,875$0.311824
WindowsC2110.9288,332271,1645,048$0.271719
WindowsC3115.8357,968336,5286,806$0.331923
WindowsC4105.6349,957329,8896,637$0.331921
Claude Code, Chrome DevTools, MCP
macOSC164.2518,779482,7365,147$0.381813
macOSC284.6514,662476,9085,675$0.401814
macOSC366.3467,466429,7595,553$0.391615
macOSC469.8454,948418,3935,455$0.381614
WSLC161.9461,992425,8695,078$0.371614
WSLC283.8590,591557,4516,481$0.412218
WSLC369.2535,587495,2445,490$0.411815
WSLC4106.8559,345526,1976,512$0.402116
WindowsC178.8592,301555,7086,171$0.422014
WindowsC2111.9482,381446,0175,828$0.391715
WindowsC369.0488,951450,9305,696$0.391717
WindowsC473.2528,741488,7495,632$0.411816
Codex, agent-browser, CLI + skill
macOSC190.1502,096467,4562,210$1.001627
macOSC295.3358,084327,4242,719$0.801527
macOSC394.4427,258384,0002,625$0.951427
macOSC483.5478,797446,7202,286$0.961627
WSLC178.8394,667359,2962,313$0.861425
WSLC270.0278,405254,0802,046$0.681223
WSLC375.4396,641366,3362,190$0.851423
WSLC4126.8424,608390,7842,105$0.911423
WindowsC192.3419,482376,4481,802$0.901326
WindowsC2106.2437,748396,9282,126$0.911429
WindowsC3104.8478,086436,7361,856$0.941529
WindowsC490.6505,292460,4161,616$0.991528
Codex, agent-browser, CLI, no skill
macOSC190.5290,195268,8002,401$0.681422
macOSC267.6281,581261,7601,857$0.631424
macOSC376.8334,785315,1361,883$0.681722
macOSC475.9317,529297,0881,875$0.671625
WSLC159.3227,881203,9041,904$0.541320
WSLC278.2287,027232,7041,905$0.871419
WSLC359.2265,128215,0401,641$0.801418
WSLC468.6256,473235,9041,871$0.611320
WindowsC190.0291,446259,7122,273$0.691423
WindowsC287.1272,835254,3361,867$0.611423
WindowsC3107.0318,932297,6002,269$0.701627
WindowsC495.3308,420268,8001,644$0.751524
Codex, agent-browser, MCP
macOSC186.3473,068438,2722,099$0.981421
macOSC276.4411,651364,1602,080$0.971323
macOSC377.0408,954376,5761,839$0.871320
macOSC466.8417,015384,0001,979$0.891320
WSLC167.8336,157304,7681,821$0.791121
WSLC260.9334,115303,2321,755$0.781120
WSLC362.4341,903308,0961,790$0.811120
WSLC463.5332,658301,8241,636$0.771118
WindowsC181.7425,217391,4241,838$0.901322
WindowsC277.2429,452393,8561,799$0.921321
WindowsC380.2463,736429,8242,022$0.951420
WindowsC459.2417,455385,7921,778$0.881322
Codex, Playwright, CLI + skill
macOSC1106.4393,981362,4962,526$0.881820
macOSC291.6282,923241,5362,243$0.841520
macOSC393.9329,536309,2482,227$0.701619
macOSC470.2331,614311,4241,903$0.681617
WSLC1103.9330,372289,9202,074$0.871616
WSLC299.3323,012280,4482,672$0.921615
WSLC390.1312,307269,3122,176$0.881619
WSLC478.2210,642183,2961,975$0.631216
WindowsC1137.1290,886263,8082,209$0.651414
WindowsC2150.7399,080370,9442,227$0.761914
WindowsC3126.6306,830258,9441,801$0.831515
WindowsC4128.3291,876273,5361,911$0.631415
Codex, Playwright, CLI, no skill
macOSC184.3223,557205,0562,130$0.571217
macOSC273.9233,909206,7202,003$0.651416
macOSC397.2247,066230,1442,337$0.591515
macOSC4105.9293,966257,5362,239$0.811719
WSLC181.5246,482230,2722,388$0.591418
WSLC280.7223,968193,4082,097$0.681421
WSLC379.1230,884196,8642,027$0.671418
WSLC481.9239,769223,6162,171$0.571414
WindowsC1136.0306,016278,0162,518$0.681610
WindowsC2124.9209,276185,0882,304$0.621313
WindowsC3160.3367,900332,5442,399$0.882020
WindowsC4152.6333,621314,6242,174$0.691821
Codex, Playwright, MCP
macOSC1116.5334,699310,7842,215$0.741518
macOSC2113.1300,783274,8162,066$0.711615
macOSC371.7281,815264,4481,609$0.591519
macOSC466.1241,582223,8721,848$0.571317
WSLC189.4283,805260,4802,095$0.631519
WSLC279.1244,789221,3121,975$0.581318
WSLC383.6239,171211,7122,113$0.671317
WSLC470.1265,970247,2961,984$0.611419
WindowsC196.6227,696200,8322,198$0.581216
WindowsC2111.2276,835257,4082,512$0.661414
WindowsC389.6255,920234,4962,071$0.631219
WindowsC487.3188,200160,1282,189$0.551014
Codex, Chrome DevTools, CLI + skill
macOSC167.9186,379169,0881,949$0.521016
macOSC268.9200,361178,1761,807$0.521117
macOSC373.2184,909160,0001,820$0.531017
macOSC470.4220,817199,0401,967$0.551217
WSLC176.0217,837183,9361,820$0.691217
WSLC274.4194,059180,7361,849$0.481117
WSLC378.4257,037238,8481,956$0.591417
WSLC472.1193,834177,4081,837$0.511117
WindowsC1139.9232,539206,5921,916$0.561217
WindowsC2142.5172,584150,2721,708$0.461016
WindowsC3141.9189,391171,3922,007$0.531017
WindowsC4167.8222,617207,8721,925$0.531317
Codex, Chrome DevTools, CLI, no skill
macOSC173.5207,489190,7201,804$0.521217
macOSC287.4283,141264,3202,120$0.631517
macOSC385.2234,010213,1201,891$0.591317
macOSC470.7226,207198,5281,795$0.641317
WSLC1107.2233,125215,8081,996$0.561318
WSLC268.8227,191210,4321,855$0.551318
WSLC379.5232,221214,6561,953$0.561318
WSLC473.5232,059214,7841,860$0.561318
WindowsC1144.7239,154212,6081,965$0.581318
WindowsC2138.1239,063202,3681,910$0.661317
WindowsC3168.5263,123214,0162,111$0.811418
WindowsC4276.5231,905190,0802,003$0.801318
Codex, Chrome DevTools, MCP
macOSC159.9165,715136,9601,956$0.601114
macOSC281.7339,298298,4962,113$0.891715
macOSC364.1212,156195,5841,931$0.551114
macOSC488.5209,419193,5362,075$0.531215
WSLC180.6257,717236,2882,079$0.631714
WSLC262.5247,106221,5681,839$0.641214
WSLC366.5173,341151,2961,872$0.471114
WSLC466.2247,873228,7361,854$0.591215
WindowsC193.0365,295336,6401,748$0.711714
WindowsC2112.2241,473215,2962,122$0.581215
WindowsC388.3233,948213,2482,093$0.601115
WindowsC484.2189,917158,0802,064$0.661115

Source: my study runs, 19 to 22 October 2026.

Cite this study

Maximilian Fabian Kirchner (2026). Testing agent-browser, Playwright and Chrome DevTools with Claude Code and Codex. maximilianfabian.dev, published 1 November 2026. https://maximilianfabian.dev/blog/browser-tools-for-agents/

@misc{kirchner2026-browser-tools-for-agents,
author = {Kirchner, Maximilian Fabian},
title = {{Testing agent-browser, Playwright and Chrome DevTools with Claude Code and Codex}},
howpublished = {\url{https://maximilianfabian.dev/blog/browser-tools-for-agents/}},
year = {2026},
month = nov,
note = {Published 1 November 2026},
}

To hand this study to an agent, give it the Markdown copy, which holds its text, its tables and each chart as text. The site’s llms.txt lists the copy of every post and guide.