agent-browser: Rust-native browser automation CLI for AI agents

My notes

Summary

agent-browser is a Rust-native CLI (no Node or Playwright runtime needed) that drives Chrome for AI agents through an accessibility-tree snapshot-and-ref workflow: snapshot returns refs like @e1 and @e2, and click or fill then act on those refs. It ships as a single fast binary with a persistent background daemon, an optional MCP server, a local live-viewport dashboard, visual and snapshot diffing, React DevTools introspection, and pluggable cloud browser providers (Browserbase, Browserless, Kernel, AgentCore, Browser Use) as drop-in fallbacks when a local headless browser is not viable or gets bot-detected.

Key Insight

• What the Rust rewrite bought (reported figures, not published in the repo): a widely shared TikTok breakdown of the rewrite claimed the install shrank from 710 MB to 7 MB, using roughly 18x less memory, with a 1.6x faster cold start and 93% fewer tokens from snapshot versus dumping the full DOM. The repo itself publishes none of these numbers, so treat the exact figures as unverified. The direction is consistent with the architecture though: dropping the Node runtime and Playwright’s bundled browser binaries leaves a single binary that speaks CDP to Chrome directly, and a compressed accessibility-tree snapshot is inherently cheaper to feed a model than a full DOM.
• Security is opt-in but CLI-native, not prompt-based: --confirm-actions eval,download and --action-policy <path> gate specific action categories at the tool level, so approval logic lives in the daemon rather than in an agent’s self-discipline. --allowed-domains also blocks sub-resource and WebSocket traffic to non-allowlisted domains, not just top-level navigation.
• diff is a first-class subcommand, not a manual screenshot-and-eyeball step: diff snapshot (structural accessibility-tree diff), diff screenshot --baseline before.png -t 0.2 (pixel diff with an adjustable color threshold), and diff url <urlA> <urlB> --screenshot --selector "#main" (compare two live URLs directly, with no manual before/after capture).
• dashboard start runs a standalone background server (port 4848, independent of any browser session) with a live JPEG viewport stream, a chronological activity and command feed, and console output. It lets a human watch an agent mid-task instead of inferring state from text snapshots.
• The default 25s per-action timeout is intentionally shorter than the CLI’s 30s IPC read timeout (AGENT_BROWSER_DEFAULT_TIMEOUT, default 25000). Raising it above 30000 does not simply allow slower pages, it causes EAGAIN errors because the CLI’s own read timeout expires first while the daemon is still working.
• The CLI serves its own version-matched documentation: agent-browser skills get core prints instructions baked into the installed binary, explicitly so that agent instructions never drift from the installed CLI version between frequent releases. Static cheat sheets go stale by design.
• Cloud provider swap is a one-flag escalation, not a rewrite: -p browserbase|browserless|browseruse|kernel|agentcore plus an API key env var replaces the local Chrome daemon with a remote session, and every other command stays identical. Useful as a fallback rung between a local headless Chrome and giving up when a site starts detecting the local browser.
• React DevTools and Web Vitals are built in: --enable react-devtools at launch unlocks react tree, react inspect <fiberId>, react renders start/stop, and react suspense, while vitals reports LCP, CLS, TTFB, FCP and INP and works on any framework without the flag. That turns the CLI into a lightweight frontend performance and debugging tool, not just a form-filler.
• Refs are deterministic and reusable across a snapshot’s lifetime but go stale on navigation or DOM mutation. A blocked click reports the exact covering element (for example, covered by a consent-banner div), which is enough to dismiss it and re-snapshot without guessing.