Skip to content

Browser Use 1.0.8rc1 Fixes Judge Billing and Headless Login

Browser Use has published a release candidate for 1.0.8 that changes how the judge step is billed and lets runs be approved without an API key on headless machines.

Draftship Editorial4 min read

Browser Use 1.0.8rc1 arrives as a pre-release

Browser Use has published the browser-use 1.0.8rc1 tag, a release candidate for version 1.0.8 of its browser automation library, released on 20 November. GitHub labels the tag a pre-release, and the notes call it a release candidate for 1.0.8, so this is not the finished version. The 1.0.8rc1 release notes list six changes tied to numbered pull requests, with #6489 as the release PR. Install it with pip install 'browser-use==1.0.8rc1'.

Browser Use drives a real browser with an AI model. A run typically ends with a judge step, an extra model call that inspects what the agent did and returns a verdict. Several changes in this candidate touch that judge call, the approval step at the start of a run, and what the command line prints.

The six changes, as listed

  • Judge served from cache. The judge asks as one more turn of the run, so the provider reads the run from its cache instead of billing the trace again. The notes say it returns the same verdict fields, including the gotchas (#6457).
  • Approval on a headless box. No API key over SSH or on a headless machine: the run prints a link and a code to approve on any device, then continues (#6427).
  • DeepSeek and Qwen in thinking mode. When the model refuses the forced final call at the step limit, the agent asks again without forcing, so the run returns its answer (#6453).
  • Password fields stay out of output. bu.type never prints a password field's value (#6455).
  • CDP attach keeps the open page. Browser(cdp_url=...) attaches the one page already open instead of a new blank tab (#6456).
  • Smaller credit reservation. Browser Use gateway calls reserve credit for 16,384 output tokens instead of 32,768, so small balances are not refused mid-run (#6461).

How the judge and approval changes are described

The judge change is explained in terms of turns. The judge is not a separate job but another turn of the same run, so the provider can answer it from the cache it already holds for that run rather than billing the trace a second time. The notes say the returned fields are the same, which suggests code that reads a verdict should not need to change.

The approval change targets machines where nobody can open a browser. Instead of needing an API key on the box, the run prints a link and a code, you approve on any device, and the run continues. The notes do not describe the protocol behind that handshake or how long the code stays valid.

What this likely means for people building with agents

Judge caching suggests lower and more predictable cost for repeated checks within a run, which matters when an agent loops through many steps. Headless approval likely removes a common blocker for agents running in CI jobs, scheduled tasks or containers, where an interactive browser login is not an option. The DeepSeek and Qwen change suggests fewer runs that end with no answer when the model refuses a forced final call. The password fix reduces the chance that credentials show up in terminal output or captured logs. The CDP change matters if you attach to a browser session you already control and expect to keep the page in front of you.

Limits and open questions

  • The notes do not say which model providers the judge cache applies to, or whether a cached verdict can be reused across separate runs.
  • There is no detail on the approval handshake beyond the link and the code, and no statement on which providers or gateways support it.
  • No date is given for the final 1.0.8 release.
  • The credit change is described only as a reservation figure for gateway calls. The notes do not quantify what it saves in practice.
  • Because this is a pre-release, the notes give an explicit pinned install command. Whether an unpinned install picks it up is not stated.

What it means for building apps with AI

Browser automation is one way to give an app hands: a model that reads a page and clicks through it. That is a dependency you add to your own code, and this candidate is about the internals of running such an agent safely and cheaply rather than a new surface to build on. If you are adding model calls to an app, the guide to adding AI to your app covers that side of the work. Draftship builds and hosts apps from a plain language description, with a built-in database, sign-in and hosting; a browser agent remains something you wire in yourself.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.