What it is
agent-browser is a browser automation library that feeds AI agents compressed page representations instead of full DOM trees. The token-efficiency claim β 200-400 tokens per page versus 3,000-5,000 for raw DOM β is what independent testers consistently verify in early usage reports. Built as a non-AI infrastructure tool for developers writing autonomous agents. The audience skewing toward it is AI/ML engineers and backend developers who hit context-window limits with Playwright or Selenium when chaining multi-step browser tasks.
At a glance
This tool solves a real architectural problem in AI browser automation. Traditional tools like Playwright send the entire DOM (3000-5000 tokens) to the LLM on every interaction. agent-browser uses snapshot-based refs that consume 200-400 tokens instead β a 90% reduction multiple Reddit developers confirmed in practice. It also freezes JavaScript state during snapshots, eliminating the timing bugs that plague other automation tools when elements load mid-interaction.
Strong evidenceQuality score
agent-browser Browser automation for AI agents with compact state snapshots, but needs operator supervision.
This score is our editorial judgment, computed automatically from the sources, weights, and dates shown above. It reflects the data we could verify as of November 12, 2026, not a guarantee or statement of fact about agent-browser. Third-party ratings and quotes belong to their original platforms and authors. Thin data lowers our confidence label, and we say so instead of guessing. Work on agent-browser? Dispute any datapoint and we will review it, publish your response, and correct verified errors.
Individual plan details haven't been verified yet β they'll appear here on the next data refresh.
Community feedback
Ratings and quoted comments below are aggregated from third-party sources and reflect those users' views, not SearchTools.ai's.
themes inside the Sentiment pillar β not score ingredients
βTL;DR: Vercel released agent-browser, a CLI for AI browser automation that uses snapshot-based refs instead of DOM selectors. Claims 90% token reduction vs Playwright MCP. Tested it, the difference is real. alright so vercel dropped agent-browser yesterday and I've been testing it with claude code. the pitch: browser automation CLI designed for AI agents. uses snapshot refs instead of DOM selectors, supposedly 90% less tokens than playwright mcp. here's the actual workflow: agent-browser open exβ
βI started Agent Browser Protocol (ABP) as a challenge project in January to see if I could build an agent centric browser and capture the top score on Online Mind2Web Benchmark. I completed this goal last week and held the top score of 90.53% for all of 2 days until GPT-5.4 bet it with 92.8%. My main insight on an agentic centric browser is that agents are really good at turn based chat and bad at continuous time decision making. To max out LLMs on browser use, I needed to turn browsing into mulβ
βpretty wild that a single dev could fork Chrome and ship something like this in a couple months with AI assistance. The Mind2Web score is impressive too.Definitely going to try plugging this into Claude Code and see how it behaves on some messy real-world sites.β
βThe freezing JS thing is the real insight here. I've been running browser automation stuff for a couple months and the biggest failure mode is always the agent clicking on elements mid-transition or before they've fully loaded. Most setups just grab a snapshot and pray the DOM is stable. How does ABP handle heavy SPA pages where content lazy-loads on scroll? That's where basically everything breaks for me. 90% on Mind2Web solo is wild. GPT-5.4 passing it in 2 days just shows how fast the modβ
Watch & learn

Vercel Agent Desktop + Agent Browser Let My AI Leave the Browser | Claude Code Β· Claude in Chrome
lindy-operation18 days ago

This CLI Lets AI Agents Control Your Browser (Thanks to Vercel)
CandDev15 days ago

agent-browser GitHub Explained: Visual Regression Beyond DOM Checks
alexander-hitt1 month ago

Agent Browser vs Playwright MCP: Can Vercel's CLI Cut Your AI Agent's Browser Tokens?
digestible-ai15 days ago

Episode 1: What Is agent-browser? AI Browser Automation Tool π
CodenboxAutomationLab19 days ago

Agent Browser Explained: Is It Right for You? (AI Browser Automation, Fit Check)
TechOnCue14 days ago
Capabilities
Provides utilities that help programmers build, test, and ship software faster
Extracts structured data from websites and collects it for your use
Generates and runs software tests to catch bugs before code ships
The honest take
Distinct themes surfaced across user reviews β each grounded in real review text, ranked by how often it comes up.
Questions
agent-browser is a browser automation CLI tool specifically designed for AI agents that need to interact with web applications efficiently. It uses a ref-based accessibility tree approach that outputs compact text representations consuming only 200-400 tokens per snapshot, compared to 3,000-5,000 tokens for traditional DOM structures. Built in native Rust with a client-daemon architecture, it communicates with Chrome via the Chrome DevTools Protocol and provides over 50 commands for navigation, interaction, and debugging.
When you run the snapshot command, agent-browser returns a compact accessibility tree where each interactive element receives a unique reference identifier like @e1 or @e2. AI agents can then interact with specific elements by referencing these identifiers in subsequent commands, such as 'click @e2' or 'type @e5 search query'. This approach provides deterministic element targeting, faster execution without DOM re-querying, and natural text output that language models can easily parse.
agent-browser provides native binaries for macOS (ARM64 and x64), Linux (ARM64 and x64), and Windows (x64). It can be installed globally via npm or Homebrew on macOS, making it accessible across all major operating systems.
Traditional browser automation tools output full DOM structures that consume 3,000-5,000 tokens per snapshot, which quickly exhausts context windows and increases API costs. agent-browser's accessibility tree approach with ref identifiers reduces this to just 200-400 tokens per snapshot, making it significantly more efficient for AI agents that need to maintain context across multiple interactions.
agent-browser integrates with major AI development tools including Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Google Gemini, and opencode. It works seamlessly with any agent capable of executing shell commands, making it broadly compatible with existing AI development workflows.
agent-browser uses a client-daemon architecture where commands are executed through the CLI, which communicates with a persistent daemon that manages Chrome. The daemon starts automatically and persists between commands to eliminate startup overhead for sequential operations. By default, it saves configured restore state and exits after one hour of inactivity, though this timeout can be adjusted or disabled via the --idle-timeout flag.
agent-browser includes built-in observability features such as video recording, streaming, debugging tools, a profiler, and diffing capabilities. These features help developers monitor and troubleshoot their browser automation workflows, making it easier to identify issues and optimize performance.
Yes, agent-browser supports stateful workflows through sessions, profiles, authentication state persistence, cookies, storage management, proxy configuration, and security controls. This enables long-running agents that maintain context across multiple interactions, making it suitable for complex automation tasks that require persistent state.
agent-browser includes over 50 commands covering navigation, form interaction, screenshots, network control, storage management, file handling, tabs, frames, and debugging operations. It also provides first-class support for modern web applications with dedicated documentation for network control, React and Web Vitals monitoring, initialization scripts, and Next.js and Vercel deployments.
The provided information does not specify pricing details for agent-browser. You would need to visit the official website at agent-browser.dev or check the GitHub repository for current pricing and licensing information.
More Like This