Vercel Labs' agent-browser Has 49.7k Stars and a 97% Token-Saving Claim That Its Own README Does Not Make

Vercel Labs' agent-browser, a command-line tool that lets coding agents drive Chrome, gained 712 GitHub stars on 11 November and stood at 49.7k stars and 3.5k forks. The project's README does not say how many tokens it saves. A May tutorial and an open benchmark do, and they do not agree on how much that matters.

What the README says it is

The README describes "browser automation for AI agents: open pages, read them as an accessibility tree, and act on elements from the shell." It is licensed Apache-2.0 and installs with npm install -g agent-browser followed by agent-browser install, which downloads Chrome for Testing. Homebrew and npx agent-browser open example.com also work. Native binaries ship for macOS on ARM64 and x64, Linux on ARM64 and x64 with glibc and musl, and Windows on x64.

The interaction model is the selling point. "The browser stays open between commands, so each step is one short shell call." The README also promises that "a click fails, instead of reporting success, when the target is hidden, disabled, outside the viewport", which is the failure mode that makes agents report success on pages they did not change. It links a security document covering an auth vault, domain allowlists, action policies and content boundaries; the page we read did not reproduce its contents, and the direct link we tried returned a 404, so this piece cannot describe those controls.

Where the 97% figure comes from

Emmanuel John's LogRocket tutorial, published on 4 May, writes that "according to the Vercel Labs team, it can reduce token usage by up to 97% per request." The article does not describe how that was measured and does not verify it. It explains the mechanism: agent-browser snapshot -i returns only interactive elements as short references such as @E_1, so the agent never receives raw HTML.

The README we read contains no token percentage. A figure that sits in a third party's article, attributed to the vendor, is vendor-supplied for our purposes.

An open benchmark, and what it leaves out

Nicola Sosio's MIT-licensed web-agent-bench compares Playwright MCP, agent-browser, browser-use and Stagehand. The README describes it as "a fair, reproducible benchmark." It used claude-sonnet-4-6, 5 repetitions per scenario and about 300 tasks across six suites: controlled, scroll, dynamic, hard, auth and vision. Drivers are restricted to UI actions, with shell, in-page JavaScript and direct network access blocked.

Measure	Result on the benchmark page
agent-browser tokens per step, minimal API loop	about 2.7k
agent-browser tokens per step, inside Claude Code	about 35k
Pass rate, every driver	95 to 100%
Playwright MCP tokens per step	no figure given

The benchmark says Playwright MCP sends the full accessibility snapshot plus tool schemas on every call, and that most heavy-driver volume is cache reads, so the billed cost gap is smaller than the token gap. With reliability a near tie, the README's own conclusion leaves tokens, and the cost they imply, as the only difference between drivers. The run date is not given on the page, and the results are plots rather than tables.

Two descriptions of the architecture

The README calls agent-browser a native Rust CLI whose background daemon controls Chrome over the Chrome DevTools Protocol; building from source needs Node.js 24+, pnpm 11+ and Rust. LogRocket says it "uses Playwright under the hood" and mentions neither Rust nor a daemon. The tutorial is dated 4 May and the README was read on 11 November, so the two may describe different versions. The repo page showed no releases to settle it.

LogRocket also lists limits: it is primarily for Chromium-based browsers, it is not recommended for performance profiling or memory-leak analysis, and for simple static scraping plain fetch with cheerio "could be faster."

What a star count does not show

Stars are not installs. The repository showed 827 commits on main and no tagged releases, and the README's security section was a link rather than text. Chrome's Modern Web Guidance and the agent skills repos we covered earlier are the same kind of project: tooling written for agents, spreading faster than anyone has measured what it does to their cost. The number to look for is a Playwright MCP result from web-agent-bench's next published run.