Browse.sh

Product Introduction

1. Definition: Browse.sh is an open-source command-line interface (CLI) and skill catalog for browser automation, specifically engineered for AI agents. It serves as a decentralized platform for discovering, sharing, and executing standardized browser automation "recipes" (SKILL.md files) that enable AI agents to perform complex, multi-step tasks on virtually any website.
2. Core Value Proposition: Browse.sh provides a reusable, community-driven ecosystem for web automation skills, drastically reducing the development overhead and token costs for AI agents. Its primary goal is to democratize browser automation by offering a curated open web catalog of pre-built, tested, and optimized skills, allowing AI models to interact with websites as effectively as humans.

Main Features

1. Open Skill Catalog & SKILL.md Recipe System: Browse.sh operates as an open directory of browser automation skills. Each skill is encapsulated in a SKILL.md file, a declarative manifest that defines the skill's purpose, prerequisites, input parameters, and a sequence of browser primitives (clicks, types, scrolls) or API calls. This format is AI-agent-friendly, making skills interoperable across different platforms. The CLI command browse skills add [domain] installs a skill's SKILL.md locally, making it instantly available to an AI agent. Examples include browse skills add alltrails.com for trail searching or browse skills add recreation.gov for campsite availability.
2. Universal Browser Primitives & AI-Driven Control: The browse CLI exposes low-level, deterministic browser primitives that can be chained together to automate interactions. Commands like browse click "selector", browse type "text", browse scroll, and browse select allow precise control over page elements. Crucially, it supports AI-agent-native addressing, including the use of accessibility refs (e.g., browse select @8 "option"), which aligns with how modern AI models perceive DOM structure. This facilitates complex, dynamic task execution driven by natural language instructions.
3. Integrated Debugging & Real-Time Monitoring: A key feature for development and troubleshooting is the built-in network and console tailing. Running browse network --tail or browse console --tail provides a live stream of HTTP requests/responses and JavaScript logs from the automation session. This gives AI agents (and human developers) full visibility into page behavior, aiding in diagnosing failed interactions, identifying required API endpoints (XHR/Fetch), and optimizing automation flows for reliability and speed.
4. Hybrid Execution: Local & Cloud Sessions: All CLI commands function with a local Chromium instance. However, Browse.sh is designed for scale and security with its cloud integration via Browserbase. Prefixing commands with cloud (e.g., browse cloud sessions create) routes the automation through a remote, managed Chromium session on the Browserbase platform. This provides features like residential proxies, automatic CAPTCHA solving, and verified browser environments, which are essential for scraping protected sites or running unattended at scale.

Problems Solved

1. Pain Point: High Cost and Complexity of Web Automation for AI: Traditionally, enabling an AI agent to automate a website required either extensive manual coding of scripts or consuming large amounts of tokens to have the agent "figure out" the interaction in real-time. This is inefficient, error-prone, and expensive. Browse.sh solves this by providing a pre-built, optimized skill library, cutting token costs by up to 50x through suggested selectors and XHR requests.
2. Target Audience: The primary users are AI/ML developers building agent-based systems, automation engineers, technical marketers conducting competitive analysis or SEO research, and data analysts needing structured data extraction from dynamic websites. It is for anyone seeking to automate browser workflows at scale with reduced friction.
3. Use Cases: Essential scenarios include: Automated travel planning (combining skills for alltrails.com, recreation.gov, and weather.gov), e-commerce price monitoring and comparison (using skills for amazon.com, ebay.com), recruitment and job application workflows (using skills for greenhouse.com, indeed.com), government data extraction (for sam.gov, data.sfgov.org), and personal productivity (like tracking packages via fedex.com or finding restaurant reservations via opentable.com).

Unique Advantages

1. Differentiation from Competitors & Traditional Methods: Unlike traditional browser automation tools (like Selenium or Puppeteer scripts) which require bespoke code per site, Browse.sh offers a centralized, reusable skill ecosystem. Compared to using a raw LLM for browsing, which is slow and costly, Browse.sh provides pre-validated, deterministic workflows. It is not a monolithic automation tool but a modular skill repository that enhances AI agents, making it complementary to frameworks like LangChain or AutoGPT.
2. Key Innovation: The core innovation is the standardized SKILL.md format combined with an AI-agent-native CLI. This creates a two-way interface: humans and other tools can install and share skills, while AI agents can directly execute and chain these skills. The format acts as a bridge between high-level natural language goals and low-level browser actions. Furthermore, the integration of Browserbase's cloud infrastructure provides enterprise-grade execution environment management, abstracting away the complexities of browser lifecycle, proxy rotation, and anti-bot handling.

Frequently Asked Questions (FAQ)

1. How does Browse.sh reduce token costs for AI agents? Browse.sh reduces token consumption by providing pre-defined, optimized selectors and suggested XHR request endpoints within each SKILL.md file. Instead of the AI agent having to analyze the entire DOM tree to find an element (a token-heavy process), it can directly use the suggested, reliable selector. This targeted interaction model drastically minimizes the amount of page content the agent needs to process, leading to savings of up to 50x.
2. What websites are supported, and how can I add a new one? Browse.sh supports automation for virtually any website through its skill system. The open catalog includes hundreds of pre-built skills for domains like Amazon, Google Flights, LinkedIn, and government sites. Users can create and share new skills by authoring a SKILL.md file that defines the automation workflow using the standard primitives, then contribute it to the catalog or distribute it privately.
3. Is Browse.sh only for developers, or can non-technical users benefit? While authoring new SKILL.md files requires technical understanding of web structure, using existing skills is designed for simplicity and AI-agent interaction. Non-technical users benefit indirectly by leveraging AI applications (like advanced chatbots or assistants) that have Browse.sh skills installed, enabling them to perform complex web tasks via natural language. The CLI itself is developer-oriented, but its outputs (automated actions) are for everyone.
4. How does the cloud execution via Browserbase work, and why is it important? When you prefix a command with cloud, the Browse.sh CLI establishes a secure connection to Browserbase's platform, which hosts and manages a persistent Chromium browser session. This is crucial for tasks requiring long-running processes, handling sites with aggressive bot detection (using Browserbase's residential proxies and fingerprint management), or when you need to run multiple automations in parallel without resource contention on your local machine.
5. Can I integrate Browse.sh with my existing AI framework? Yes, Browse.sh is designed to be framework-agnostic. Its skills are invoked via CLI commands, which can be executed from any programming language (Python, Node.js, Go, etc.) or AI agent framework (LangChain, CrewAI, AutoGen). The agent simply needs the capability to execute shell commands to leverage the full power of the installed Browse.sh skills.