Y Combinator's Pomerado turns websites into APIs: 3 cents instead of $41

Most agents are learning to use the internet like people do, looking at a screen and moving a cursor. Pomerado, a startup from the Y Combinator ecosystem, proposes not pretending to be human. In its demonstration, Codex completed a task through computer use in 21 minutes at a cost of $41. With Pomerado, the result came in 8 minutes and 3 cents: about 2.6 times faster and about 1300 times cheaper. The project's code is open.

Why clicking buttons is so expensive

Computer use is simple, which makes it universal. The agent receives an image of the page, decides where to click, clicks, waits for a reaction, looks at the new frame and repeats the cycle. That lets it work with any website, even one with no programmatic interface. But every step requires a call to the model, and a complex task consists of dozens of steps. Hence the slowness, and hence a bill that reached $41 for a single task in the demo.

Meanwhile, a website's human-facing interface is essentially a shell around data and actions that can be requested directly. Classic integrations rely on exactly this: if a service has an API, an agent doesn't need to push a mouse around to get what a single request can return.

Website as API: what Pomerado promises

According to the project's description, Pomerado first works out how a specific website is built. The agent then talks to the site directly, and the site itself becomes an API or an MCP server. MCP (Model Context Protocol) is an open protocol for connecting tools and data sources to LLM agents, and many clients and development environments support it today. If a site is exposed to an agent as an MCP tool, any compatible agent can call it the same way it calls an ordinary function.

The second part of the promise concerns fragility. Integrations with sites that have no official API break with every redesign or markup change, and maintaining them eats a noticeable share of time. Pomerado claims that it adapts to site changes on its own, without manual rework.

Key parameters from the demonstration:

• Task for Codex via computer use: 21 minutes, $41.
• The same task with Pomerado: 8 minutes, $0.03 (3 cents).
• Time gain: about 2.6x.
• Cost gain: about 1300x.
• Result of the site analysis: an API or MCP interface for the agent.
• Adaptation: the tool reacts to site changes on its own.
• License and access: open source, repository Pomerado/pomerado on GitHub.
• Status: the project was spotted on the Y Combinator startup list.

What to check before adopting it

The figures come from a single demonstration, and they cannot be compared with anything beyond that specific task. The gap of 8 minutes versus 21 is understandable and plausible. But 3 cents versus $41 mostly shows how expensive it is for an agent to work step by step, calling a model on every action. On other sites and in longer scenarios, the ratio will be different.

The practical question for teams building agents is how well such a layer handles authentication, dynamic pages, and actions with side effects, such as making a payment or submitting a form. That is what you should test on your own scenarios.

The easiest way to start is with a pilot. Take two or three routine agent tasks on external sites, such as collecting data from user accounts or filling in repetitive forms. Run them through computer use and through Pomerado, then measure time, cost and the share of successful runs. Because the code is open, you can read exactly how the project analyzes sites and assess the risks before connecting it to production workflows. Separately, check the sites' terms of use: direct access that bypasses the interface is not permitted everywhere.

If the claimed figures hold up for even some of the tasks, computer use will remain a fallback for sites without a predictable structure, while the main load will shift to direct calls.