Startups & Funding

Apify approaches $100M ARR as AI agents fuel tool demand

Apify says AI-agent usage is pushing gross ARR toward $100 million, while its creator marketplace paid about $2 million last month to a few hundred creators.

Jan Čurn, Co-Founder and CEO of Apify – Interview Series
Jan Čurn, Co-Founder and CEO of Apify – Interview Series2048492 / pixabay
By Rebecca Stone6 min read

Why it matters

  • Apify was founded in 2015 by Jan Čurn and Jakub Balada through the Y Combinator Fellowship.
  • Čurn says gross ARR is quickly approaching $100 million after agent usage accelerated at the end of 2025.
  • Apify’s platform has more than 90,000 tools spanning web data, scraping, automation and AI.
  • More than 300 community Actors generate more than $1,000 per month.
  • Apify paid a few hundred creators about $2 million in total net earnings last month.

By Rebecca Stone

Apify’s gross annual recurring revenue is quickly approaching $100 million, co-founder and CEO Jan Čurn said, as AI agents increasingly use the company’s web-scraping, browser-automation and data tools.

Čurn said usage took off after models became capable enough for agents to perform meaningful work at the end of 2025. His account puts tool discovery, fresh data and permission controls at the center of agent deployment, rather than model capability alone.

What started Apify?

Čurn said Apify began in 2015 when he and Jakub Balada launched the company through the Y Combinator Fellowship. Their original product aimed to give developers flexible, scalable web crawlers.

Čurn said a client project involving new apartment listings in Prague exposed the gap. Existing tools struggled with dynamic JavaScript and irregular website structures, so the founders built a tool that could turn any website into an API.

Over time, Apify expanded from crawling into web data, browser automation and a marketplace for ready-to-run tools called Actors. Developers can publish and monetize Actors; businesses and AI builders use them to retrieve structured web data and automate tasks.

When did agents become Apify’s main growth engine?

Čurn said his first breakthrough came in early 2023, soon after ChatGPT launched. The release showed developers they could feed fresh web data into a prompt instead of training a new model.

Čurn said Apify was among the first scraping platforms to supply crawlers for emerging RAG pipelines and integrate with libraries including LangChain and LlamaIndex. He used a product-price comparison across 20 e-commerce sites to illustrate the agent problem: one agent would need data from many changing stores.

When Anthropic released the Model Context Protocol at the end of 2024, Apify adopted it immediately. Čurn said Apify already had thousands of tools with clear inputs and outputs, the format agents need to call them.

The real turning point came at the end of 2025, when Čurn said models became good enough for agents to complete meaningful work. Apify usage then accelerated, and the company quickly approached $100 million in gross ARR.

Why buy a tool instead of building one?

Čurn said Apify’s platform now has more than 90,000 tools spanning web data, scraping, automation and AI. He expects models to generate increasingly complex software from scratch, but says a cost threshold will determine when an agent should buy an existing tool.

Čurn said the threshold will rise as models improve, but some demand will remain. Pre-built tools should retain an edge when they depend on private data or context that is difficult to obtain.

Discovery, he said, is becoming the main constraint as software creation gets cheaper and the number of tools grows. Agents need to locate services that are both trusted and valuable for the task.

Čurn framed the economics with a direct comparison: “And rather than paying $500 to generate it properly, you can pay someone $5 to rent an existing service.”

Where are agents using web tools?

Čurn cited use cases that he described as unusually specific:

  • An AI system scans YouTube, eBay and Japanese shops to determine which Pokémon cards are worth buying.
  • An insurer used Apify to check 55 Italian obituary sites and verify policyholder deaths.
  • Canary uses Apify for Spotlight, a tool that helps investigators find victims of child trafficking.

Čurn said these projects share a specific problem and access to the right data. He contrasted them with agents designed to perform a bit of everything.

Why do fresh data and silent failures matter?

Čurn characterized a large language model as a lossy compression of its training data. It can make clever guesses, but it still guesses, he said, making it unreliable for real-world decisions that depend on current conditions.

Once agents make decisions or complete tasks in the real world, they cannot rely entirely on static training data, according to Čurn. Prices change, products disappear, competitors launch products and websites update constantly. Models can supply reasoning, but agents still need current web information converted into structured data at scale.

Modern websites add further failure modes. Čurn listed dynamic JavaScript, challenge pages, incomplete responses, changing interfaces, authentication requirements and anti-bot systems that inspect IP reputation and browser fingerprints.

“Failures are often silent, so an agent can confidently summarize an empty page without noticing,” Čurn said. He contrasted a system that works once with production software that must survive continuous website changes.

What must change before agents can act?

Čurn said permissions become more important when an agent can act rather than only read. His security rule was direct: the agent should never see a user’s credentials, and every secret must remain outside a secure sandbox and away from the model.

Apify is working on proxy sandboxing for Actors on its platform and in mcpc, its open-source command-line client for the Model Context Protocol. Čurn said those controls are necessary as agents move from information retrieval into online actions.

How should models, tools and browsers fit together?

Čurn expects ChatGPT and Claude to become primary interfaces through which people command computers, with external tools connecting to them as standardized services. He said tools running elsewhere connect through the Model Context Protocol, which has become the standard interaction protocol for those services.

For information such as restaurant ratings and contacts, he said, a browser can open Google Maps and click through the service. A dedicated tool that returns structured mapping data is faster and more reliable, he said.

Čurn also expects a main agent to hand off work such as booking tables and finding flights to specialized subagents running on cheaper models. He said more users are now connecting Apify into agents such as Claude and ChatGPT, while the platform and its Actors optimize for agent use to drive more usage and spending.

Can agents pay for the tools they need?

Apify’s existing marketplace already lets developers publish Actors, earn payments when other people use them and attract more tool makers. Čurn said that loop has run for years, initially with human buyers.

He expects agents eventually to carry small wallets and pay for tools during a task through protocols such as x402 or MPP. The volume remains small, he said, but he believes most economic activity will eventually occur autonomously between agents with less human oversight.

Apify paid a few hundred creators about $2 million in total net earnings last month, Čurn said. More than 300 community Actors now generate more than $1,000 per month, giving their creators recurring income.

As Apify enters its second decade, Čurn said, the company wants to become the world’s largest and most trusted marketplace for AI tools. His target extends beyond human users to what he describes as billions of future agents.

The commercial test will be whether those agent-to-agent payments and tool calls expand beyond the small volume Apify still sees.

Source: unite.ai

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

754 articles

Related articles

  1. Two Engineers Built Stack Overflow for Agents. ChatGPT Tripled Its Signups
  2. OpenAI ships a model-native harness and native sandboxes for its Agents SDK
  3. OpenAI Opens the Codex Harness to All Developers With Agents API
  4. Genesys lets companies bolt AI agents onto old contact centers
  5. Australian firm opens the books on five months of AI agent controls

« Previous article