Which agents work with it?

Any agent that reads skills from a directory. Claude Code, Codex and Gemini CLI read it natively; Cursor and VS Code connect over MCP.

The voice runs on your own machine, so there is nothing to meter. The first run downloads it, about 340 MB.

Yes. The skill is MIT-licensed and the authoring format is public.

Create

Browser Use now grades every run, and screenshot-heavy runs cost half as much

15 hours ago

@browser-use Subscribe

Browser Use now grades every run with your own model, and notes what a site rejected or blocked even when the run succeeded. Screenshot-heavy runs cost about half as much, and the CLI starts its own Chromium when it can't use your Chrome. Also fixed: dead CDP addresses, slow cancels of cloud runs, and setup errors that a newer release would solve. Covers Browser Use 1.0.3, 1.0.4 and 1.0.5. Release notes 1.0.3: https://github.com/browser-use/browser-use/releases/tag/1.0.3 Release notes 1.0.4: https://github.com/browser-use/browser-use/releases/tag/1.0.4 Release notes 1.0.5: https://github.com/browser-use/browser-use/releases/tag/1.0.5 Full changelog: https://github.com/browser-use/browser-use/compare/1.0.2...1.0.5

Ask about this presentation

Show transcript

Browser Use now grades every run. When the agent finishes, your own model makes one more quick request, checks the run against the task, and reports whether it worked, and if not, why.
Slide text
+ 01
Browser Use now grades every run with your own model
One more low-effort request checks the run against the task: did it work, why not, and what to watch for
Release notes, Nov 14 · Pull request #6398
The judge also notes what a site rejected, blocked or rate-limited, even when the run succeeded. Where a site pushed back, it now catches that every time, up from fewer than half. It checks screenshots too.
Slide text
+ 02
The judge writes down what the site pushed back on, even when the run succeeds
It sees the trace, the last screenshots and the final page, unless vision is off
“The form requires phone numbers in (###) ###-#### format”
Release notes, Nov 14 · Pull request #6403
Screenshot-heavy runs now cost about half as much. Dropping one old screenshot per step broke the provider's prompt cache. Now they go in one batch past twenty, and the agent scores the same.
Slide text
+ 03
Screenshot-heavy runs now cost about half as much
+ Old screenshots go in one batch past 20, so the prompt cache holds
+ Model cost per task: $0.21, down from $0.43, at the same score
+ The agent now keeps up to 20 old screenshots in view, not 2
Release notes, Nov 13 · Pull request #6394
The command line now brings its own browser, like MCP. With no display, or remote debugging off, it starts Chromium and keeps it across calls until you reload. Named sessions get their own.
Slide text
+ 04
The CLI now starts its own Chromium when it can't use your Chrome
+ With no display, or no browser with remote debugging on
+ It stays open across calls; browser-use --reload closes it
+ Callers that set BU_NAME get their own Chromium too
Release notes, Nov 12 · Release notes, Nov 13
Three more fixes. A CDP address with nothing behind it fails in three seconds, not thirty. A cancelled cloud run frees the browser within ten seconds. And setup errors now say when upgrading would help.
Slide text
+ 05
+ A BU_CDP_URL with nothing listening fails in 3 s, not 30, and names both ways out
+ A cancelled cloud run lets go of the browser within 10 s, not 300; downloads still copying are dropped
+ When a newer release may fix a setup error, the CLI and MCP tell the agent to upgrade
Release notes, Nov 12 · Release notes, Nov 13
Once you upgrade, the judge runs on every task. Its tokens count toward your bill. The release notes on GitHub show how to give it your own criteria, or switch it off.
Slide text
+ 06
Upgrade, and the judge runs on every task
Its tokens count toward each run's cost
Release notes on GitHub
browser-use/browser-use · 1.0.5
Release notes, Nov 14 · Pull request #6398