Search EN / ES
AI

Vercel’s Agent-Browser Replaces Playwright for AI Agents—93% Fewer Tokens

Playwright was designed for human-written tests, not AI agents, leading to slow, expensive workflows that dump full-page screenshots into context windows. Vercel’s agent-browser solves this by feeding models compact accessibility trees instead of pixels, reducing token usage by 93% and accelerating execution. The tool is already a GitHub favorite, with over 31,000 stars, and integrates seamlessly with AI coding assistants like Claude Code.

Lucas M (AI-assisted) 3 min read EN

Vercel’s agent-browser is a Rust-based CLI designed specifically for AI agents, replacing Playwright’s screenshot-heavy approach with a lightweight accessibility tree that cuts token costs by 93% and speeds up automation.

Why Playwright Falls Short for AI Agents

Playwright was built for human developers writing test scripts, not for AI-driven workflows. When AI agents use Playwright (or its MCP variant), each step typically involves:

  • Capturing a full-page screenshot
  • Sending the image to the model
  • Waiting for the model to interpret the pixels and decide the next action

This process repeats for every interaction—clicks, form submissions, or page navigations—resulting in tens of thousands of wasted tokens. The model often misses details because it’s parsing an image rather than structured data, making the workflow slow, expensive, and unreliable.

How Agent-Browser Works

Agent-browser replaces screenshots with a compact accessibility tree that labels DOM elements with references like @e1: button "Sign in". The AI model selects a reference, and the tool executes the action directly. This approach eliminates unnecessary token consumption and speeds up execution.

Key features:

  • Accessibility tree with refs: Instead of 2MB PNGs, the model receives structured text like @e1: button "Submit", reducing token usage by 93%.
  • Rust-based performance: No Node.js overhead or Playwright runtime. The tool connects directly to a Chrome instance for fast execution.
  • Semantic locators: Supports plain-language commands like agent-browser find role button click --name "Submit".
  • Screenshots on demand: Only captures pixels when explicitly needed, further reducing token waste.

Installation and Setup

Agent-browser can be installed globally in seconds. The fastest method is to use Claude Code with the prompt:

install agent-browser globally and run agent-browser install to download Chrome

For manual installation, run:

npm install -g agent-browser
agent-browser install

The second command downloads Chrome for Testing, Google’s official automation build. Mac users can also use Homebrew:

brew install agent-browser && agent-browser install

When to Use It

Agent-browser is the default choice for most AI-driven browser automation tasks, except for one-off screenshot jobs where token cost isn’t a concern. It’s particularly useful for:

  • Multi-step workflows (e.g., form submissions, data extraction)
  • Projects where token efficiency matters
  • Integrations with AI coding assistants like Claude Code

Tradeoffs

While agent-browser is optimized for AI agents, it lacks some of Playwright’s features for human-written tests, such as:

  • Detailed debugging tools for manual testers
  • Support for non-Chrome browsers (e.g., Firefox, Safari)
  • Advanced screenshot customization

For teams already invested in Playwright’s ecosystem, migrating to agent-browser may require rewriting existing test scripts.

Bottom Line

Agent-browser is the first tool built from the ground up for AI agents, not human testers. By replacing screenshots with structured accessibility trees, it slashes token costs and speeds up automation without sacrificing accuracy. With over 31,000 GitHub stars and backing from Vercel Labs, it’s quickly becoming the standard for AI-driven browser interactions.

Sources 5

  1. 01 @tenfoldmarc https://instagram.com/tenfoldmarc
  2. 02 agent-browser https://github.com/vercel-labs/agent-browser
  3. 03 Chrome for Testing https://developer.chrome.com/blog/chrome-for-testing
  4. 04 Claude Code The AI coding agent that pairs with this → https://claude.com/claude-code
  5. 05 Join AI Builders → https://paid.tenfoldmarketing.com
Similar Articles

More articles like this

AI 4 min

Claude Code: The Terminal-Based AI That Runs Your Business While You Sleep

Most Claude users never leave the browser tab. A smaller group has moved to Claude Code, a terminal-based interface that unlocks plugins, scheduled agents, MCPs, and project-aware files. This guide walks through installation, the four modes, slash commands, managed agents, skills, MCPs, and the two files that run an entire business. All for the same $20/month Pro plan.
AI 2 min

Cut Claude Code Costs

Claude Code is a powerful coding tool, but its token usage can quickly add up. By implementing three simple tricks, users can significantly reduce their token usage without compromising on performance. These tricks include using the Opus and Sonnet models efficiently, utilizing subagents for research and exploration, and installing the Caveman plugin. By combining these methods, users can extend their token usage limits and get more out of their Claude Code plan.
AI 3 min

Higgsfield MCP Server: Turn Claude Into a Short-Form Ad Factory in 2 Minutes

Higgsfield, a visual generation platform that wraps models like Seedance 2.0, Sora 2, Veo 3.1, Kling 3.0, and Hailuo 02 behind a single interface, shipped an MCP server on April 30, 2026. This lets Claude Desktop users generate short-form ads by simply chatting — no clicking around the Higgsfield UI. Nine curated presets (UGC, unboxing, product review, hyper motion, TV spot, and more) ship out of the box. The workflow collapses creative production from days to minutes, making it realistic for brands to ship the 30+ ad variants per month that Meta's algorithm rewards.
AI 2 min

OpenAI and PwC collaborate to reimagine the office of the CFO

OpenAI’s quiet alliance with PwC arms CFOs with autonomous agents capable of parsing GAAP filings, reconciling ERP ledgers, and triggering real-time audit flags—effectively outsourcing the "last mile" of financial close to transformer-based workflows. The deal signals a shift from point automation to full-stack orchestration, with PwC’s 6,000-strong AI task force embedding OpenAI’s Operator API into enterprise-grade control planes. AI-assisted, human-reviewed.
AI 2 min

DeepClaude Lets You Run Claude Code With DeepSeek's Brain for 17x Cheaper - Decrypt

A new cloud-based service, DeepClaude, slashes costs for running OpenAI's Claude large language model by leveraging the massively parallel architecture of DeepSeek's Brain, a custom-designed ASIC, to achieve a 17-fold reduction in computational expenses, making high-performance LLM inference accessible to a broader range of developers and enterprises. This breakthrough is poised to accelerate AI adoption across industries. The service's efficiency is attributed to its ability to optimize Claude's neural network for DeepSeek's Brain's unique hardware capabilities. AI-assisted, human-reviewed.
AI 4 min

59 Claude Prompts to Solve Real-Life Problems—Not Just ‘Productivity Hacks’

Claude’s potential is often wasted on generic queries. A curated set of 59 prompts—organized by real-world problems like finance, life admin, and creative problem-solving—helps users extract more value from the AI. The key? Treating Claude as a collaborative tool, not a search engine, and refining outputs through iterative feedback. Here’s how to use them effectively.