Gunjo · Business Intelligence for the AI Era
← Sticker Wall AGENT · DETAIL

Browser Use Browser Agent: A Multi-Million-Dollar Annual Web Automation "Shovel-Selling" Business

Workflow: Feed natural language tasks to the Agent daily, such as "scrape prices for five competing products and generate a compar

AGENT

Key Fields

FIELD STAMPS
IndustryContent / Creator Economy
RegionGlobal(全球(瑞士苏黎世起家,美国旧金山运营,服务全球开发者))
ScaleSME
ChannelOnline

🔧 Workflow

Feed natural language tasks to the Agent daily, such as "scrape prices for five competing products and generate a comparison table" or "automatically fill out this recruitment form." The Agent drives a real Chrome browser to handle clicks, typing, scrolling, multi-tab, and form interactions, outputting structured JSON or reports through DOM extraction combined with visual understanding, rather than brittle CSS selectors. Use n8n or cron to schedule tasks to run automatically every day, while humans handle final spot checks before delivery and iterative correction of failure cases.

🛠 Setup Requirements

Basic Python skills required. Use the single command `uv add browser-use` to install the MIT open-source library, then connect any LLM API (Claude, GPT, Gemini, or local Ollama). You can run your first real website automation script and get structured output within 1-3 days. When moving to production, switch to the official cloud API, where the platform handles stealth evasion, 195-country residential proxies, and CAPTCHAs, eliminating the need to build an anti-scraping system yourself.

🧰 Toolchain

  • 🔧 browser-use (MIT open-source Python/TypeScript library)
  • 🔧 Playwright browser kernel
  • 🔧 Claude, GPT, and other large model APIs
  • 🔧 Ollama local model for cost reduction
  • 🔧 n8n scheduled task orchestration
  • 🔧 Official cloud browser and hosted Agents API ($0.02/hour)

💰 Revenue

The individual path mainly sells "the price differential of automation capabilities": the official benchmark is about $17 cents/task (Hard Bench 82% success rate) and cloud browsers cost $0.02/hour, while external services such as competitor monitoring, batch resume submission for job applications, automated form filling, and multi-platform content distribution are generally priced at 3-5 times the cost. Estimated at a stable 400-600 automation tasks per month serving 8-15 small clients, the monthly revenue range is about $1,500-$3,000, which can scale to over $5,000 once established. The platform side itself has achieved an eight-figure ARR, proving extremely strong willingness to pay in this category.

💸 Cost

The open-source library is completely free. The main costs are LLM APIs and pay-as-you-go cloud browsers: the cloud API costs about $0.02 per browser hour, single-task benchmark cost is about 17 cents, and heavy monthly usage runs about $100-$300. If you switch to a local Ollama model, marginal costs are almost zero, though speed and success rates drop, making it suitable for practice and low-value batch scenarios.

⏱ Time Investment

Concentrated investment of 1-3 days during the setup phase to run the first script, followed by 1-2 hours of daily maintenance: checking previous night's scheduled task logs in the morning, spot-checking outputs, repairing workflows broken by website redesigns. Spend 2-3 hours each on weekends adding new task templates for old clients and building up a reusable prompt library. This is a classic low-human-labor marginal-cost business.

🚀 Getting Started

The first step costs nothing: clone the GitHub browser-use/browser-use repository, install it with `uv add browser-use`, connect a Claude or GPT API Key, and first run a small script to "scrape prices for five items on an e-commerce platform and output JSON." Once working, wrap the script into a scheduled task, then list services for "web automation data scraping and automated form filling" on Xianyu, Upwork, or Fiverr to land your first batch of orders, using the free open-source version for fulfillment and switching to the cloud API when anti-scraping measures are needed.

🔑 Keys to Success

  • ✅ Choose vertical must-have pain points to enter: competitor price monitoring, batch job applications, form filling, multi-platform content distribution, end-to-end regression testing. Business must be tied to human review; purely flashy demos make no money.
  • ✅ Humans act as referees: automation only does the work. Manual spot-checks are mandatory before delivery, and failure cases must be fed back to refine prompts and action sequences, building compoundable data assets and task template libraries.
  • ✅ Capitalize on platform low-cost advantages: $0.02/hour and 17 cents/task costs are far lower than OpenAI Operator, leaving ample room for external pricing. Build volume and acquire clients first before raising prices.
  • ✅ Prioritize official models and cloud APIs: The official browser-dedicated model is 3-5 times faster than general models, and combined with stealth and residential proxies, penetration rates are high, bringing single-task costs down to the 17-cent tier.

⚠️ 风险

  • ⚠️ Website anti-scraping and redesigns: Logins, CAPTCHAs, and DOM structure changes can break tasks at any time. Official stealth and 195-country residential proxies can mitigate but not completely solve this, requiring ongoing script maintenance.
  • ⚠️ Data compliance and terms of service risks: Batch scraping third-party data or automated form filling may violate target website terms of service. When personal data is involved, client written authorization is required, and sensitive fields must be proactively avoided.
  • ⚠️ LLM hallucinations leading to operational errors: The model may misunderstand pages during critical transaction steps. Actions like placing orders, making payments, and submitting official forms must include a human confirmation step to prevent accidents and costly order liabilities.
  • ⚠️ Homogeneous competition: Free open-source lowers the barrier to entry to virtually zero. Simply selling scripts will lead to a race to the bottom; you must build moats through industry knowledge subscriptions, private deployments, or long-term client relationships.

📌 Real Cases

  • 📌 Manus deeply integrated Browser Use as the underlying browser control component, directly driving its GitHub downloads from 5K to 28K+ per day. Manus itself achieved roughly $100M in ARR and was later acquired by Meta for about $2B, proving that "equipping Agents with browsers" is a genuine rigid demand that delivers commercial returns to integrators.
  • 📌 Browser Use company itself: Completed a $17M seed round in March 2025 led by Felicis with participation from YC-affiliated parties like Paul Graham. With a team of only about 7 people, it achieved an eight-figure ARR, serving as a benchmark for "selling shovels" in the Agent Infra track.
  • 📌 The official Hard Bench benchmark achieved an 82% success rate with a cost of 17 cents per task, outperforming frontier models like Opus 5 and Sonnet 5 running raw. This is reliability evidence that individual freelancers can present to enterprise clients, directly lowering customer acquisition trust costs.