How We Deploy Custom Chromium
- Source
- browserbase.com
- Author
- Browserbase
- Date

Context
Browserbase, which runs cloud browsers for AI agents, says it can now build and test its patched version of Chromium before Google ships the matching Chrome release, where an upgrade once took about a week. Its engineering post explains why that is hard. The company maintains more than 18,000 lines of its own changes to Chromium, the open-source browser behind Chrome, for security, performance and "stealth", meaning agents can browse without being flagged as bots. Google now ships a major Chromium version every two weeks, and the post counts 400 to 500 new security vulnerabilities a month in Chromium. Customers rely on Browserbase to keep the browser running their sessions up to date, so the company tries to keep the gap between an upstream fix and its own release short.
The main change was making the patches understandable enough for coding agents to carry forward. In 2024 engineers kept the list in Notion and re-applied each patch by hand, comparing old and new code side by side. Now each patch has a name, a written record of why it exists, and regression tests. Agents work in a loop: inspect a patch that no longer applies, find where the relevant code moved upstream, regenerate the change, replay the full patch series and run the tests. Internal tools also watch Chromium's public repository; a tag with real code changes followed by one that changes only the version file signals a version worth building early.
Passing tests turned out to be insufficient, so Browserbase keeps three separate checks. Patch tests and binary tests check behavior in isolation; application tests run the browser with its production configuration. The post's example is Chromium 157, which made HTTPS-First mode the default: plain HTTP pages could be retried over HTTPS and fail, but the test suite missed it because local test addresses are exempt. Only the application check caught it, and the team then added a test that disguises a local server as a public HTTP site.
Builds were the other bottleneck, once taking about six hours. Browserbase had named its cache folders after the Chromium version, so every upgrade moved the workspace, changed absolute paths in thousands of build commands and forced a near-full rebuild; a single stable folder fixed that. The team also tried a faster way to arrange code during linking, which saved about 54 seconds per link but made the browser start 3.33% slower, and it reverted that change. Rollouts go to a small share of machines first, and agents confirm that the running session reports the intended version. Rolling back is not simple, because a newer browser may already have converted user data into a format the older one cannot read.
- Browserbase gives coding agents named patches, intent documentation and exhaustive regression tests, then runs them in a loop: inspect a rejected patch, find where the behavior moved upstream, regenerate the diff, replay the series and test.
- Passing tests were not enough: Chromium 157 enabled HTTPS-First balanced mode by default, which binary tests missed because loopback hosts were exempt. Browserbase caught it in application verification and added a test mapping a public hostname on port 80 to a local server.
- Browserbase's slow builds came partly from cache directories named by Chromium version: absolute paths in build state meant over 3,586 completed steps went unrecognized after each bump. A stable workspace across versions fixed it after one rebuild.
- Switching the linker's function-ordering algorithm from cdsort to hfsort cut about 54 seconds from linking but slowed browser startup by 3.33%, so Browserbase switched back to keep customer browsers fast.
- Rolling back a Chromium upgrade is not just restoring the old executable, Browserbase notes: the new version may already have migrated the user data directory into a format the old one cannot read.
Checking sign-in…
Loading comments…






