Test results
Which browser tool helps AI the most?
We raced 7 browser tools for AI through the same game with the same AI. Playwriter won the most levels. Our first try was the fastest and the cheapest, but came 4th.
Updated October 16, 2026 · by the Buoy team
95%
of levels won by Playwriter, the top tool
4.8 s
a level for Buoy, the fastest
83k
tokens a win for Buoy, the least
The game
A web page that keeps score.
We built a game called Page Race. Each level asks for one small job, like click a button, fill a form or drag a card. Our server decides every win, so no tool times itself.
60
levels in 6 kinds
7
hard cases, scored apart
885
runs in all
1
AI model for all 7 tools
All 7 tools used Claude Haiku 5.5 at its normal effort. Only real clicks and keys count. Each tool was set up the way its own guide says.
Results
Who finished the most levels.
Finishing comes first. Speed breaks ties.
The tool drives the Chrome you already have open and signed in.
mcp-chrome clicks with page script, which the game doesn't count as real. When those clicks count too, it wins 92% of levels.
Speed
Buoy finished each level first.
This is the middle time for one level, from the AI's first step to its last. It counts the AI's thinking too, since that's what you wait for.
Only tools that won at least half the levels are shown here.
Fast hands
Click fast. Answer fast.
A tool can land a click fast but tell the AI late. The AI waits for that answer before it can act again. Here a script plays with no AI, so we time only the tool.
lands in 8 ms · answers under 1 ms later
lands in 9 ms · answers under 1 ms later
lands in 26 ms · answers 1 ms later
lands in 26 ms · answers 33 ms later
lands in 21 ms · answers 204 ms later
lands in 16 ms · answers 1010 ms later
The hardest level was the double flash. A target lights up for 0.4 s, goes dark, then lights up again. No tool hit both flashes with the AI driving. The AI takes seconds to think.
Where each tool slips
Six kinds of jobs.
Each kind counts the same. Darker means more levels won.
The tool drives the Chrome you already have open and signed in.
| Tool | Find & click | Type & forms | Wait & timing | Scroll & drag | Read & choose | Chains |
|---|---|---|---|---|---|---|
| Playwriter | 100% | 100% | 80% | 90% | 100% | 100% |
| Playwright MCP | 100% | 100% | 80% | 100% | 100% | 80% |
| Chrome DevTools MCP | 100% | 100% | 70% | 100% | 100% | 80% |
| Buoy | 100% | 100% | 70% | 70% | 100% | 90% |
| agent-browser | 100% | 90% | 80% | 80% | 80% | 70% |
| Claude in Chrome | 40% | 30% | 20% | 40% | 20% | 0% |
| mcp-chrome | 0% | 0% | 0% | 0% | 0% | 0% |
Hard cases
These are scored on their own: did the tool get the job done in our race, or not.
Ten in a row
One prompt, ten levels.
We gave the AI one job: finish ten levels back to back. Each next level shows only after the last one is won.
The tool drives the Chrome you already have open and signed in.
Most levels won
Time for all ten
Buoy lost the same two drag levels every time. The time chart shows only tools that won all ten.
Cost
What each win costs.
Some tools send the AI a big copy of the page after each step. That fills its memory and costs more.
Tokens for each level won, in thousands. Fewer is better.
Every number
The fine print.
We make Buoy and we built this game. So here is how we tested, and every number.
How we tested
- We locked every level and rule before Buoy joined the race.
- Each tool played each level once in each lane it supports, with a fresh AI each time.
- The game logs each input and checks that it is real. Fake "skip" buttons catch cheat tries.
- The agents could not read the game's code or the judge.
- We ran 6 races at once on one Mac, all tools mixed in the same order.
- The no-AI timing ran one tool at a time on a quiet Mac.
Every result
The tool drives the Chrome you already have open and signed in.
Fair notes
- Buoy's web tool is a preview. It is not out yet.
- Buoy got one cheat flag. It could only drop a card on a button, not on a plain box. The AI got stuck and dropped the card on the "Skip level" button to see what would happen. We kept the flag.
- Claude in Chrome works in its own tab group. Most times it could not reach the race tab that was already open.
- mcp-chrome clicks with page script, so its clicks don't count on the main board.
- The ten-level run ran four times for each tool in each lane. We show the best result and the middle time.
- We fixed two bugs in our no-AI timing before the final run. Both hurt fast tools the most. Then we ran it again for all of them.
- One Mac, one AI. Other AI models may rank the tools in a different order.
Tool versions
Can I use Buoy for the web?
Why does Buoy finish fast but win less?
Why one AI model?
sources
Capability claims on this page come from each vendor's own documentation, read on the date shown. We did not install and run every tool listed.
- Playwright MCP (Microsoft) · checked October 16, 2026
- Chrome DevTools MCP (Google) · checked October 16, 2026
- agent-browser (Vercel) · checked October 16, 2026
- Playwriter · checked October 16, 2026
- mcp-chrome · checked October 16, 2026
- Claude in Chrome (Anthropic) · checked October 16, 2026