Conversations
Pomerado launches self-service platform that turns websites into MCP interfaces
The Pomerado team introduced its YC F26 project on Hacker News: users describe a workflow in natural language and answer a few questions, and the platform can generate a repeatably callable MCP within an hour for websites that lack APIs. The system first demonstrates the task through a browser, then generates HTTP requests or Playwright scripts, and sets up safeguards, tests, authentication, and automatic repair mechanisms after failures.
Pomerado was launched by Alan and Akshay with the goal of letting AI agents connect to any website. The team said websites in sectors such as aviation, banking, e-commerce, insurance, government, and traditional portals often lack APIs or MCPs, leaving agents to rely on browser automation or computer operation. Such approaches are slow, costly, and unstable, and they may also be disabled by enterprises for security and compliance reasons.
According to the introduction, developers can create an API for a website on the self-service platform even if they do not own the website. Users or their coding agents only need to provide a short natural-language description and answer a few questions about the workflow, after which they can usually obtain an MCP within an hour. Subsequent tool calls are completed through the deterministic MCP, with most requests returning results within 5 to 30 seconds.
During generation, Pomerado's agent opens a browser, explores the website, and performs the target task once, writing HTTP requests or Playwright scripts and recording navigation, control interactions, and how results are obtained. If login or further clarification is needed, the system asks the user through the MCP. The agent then generates a generalizable, repeatable API based on the demonstration run and recorded data.
The platform also sets up multiple layers of validation: an independent Guardian model reviews the coding agent's calls according to safeguard rules; the system tests with synthetic inputs to prevent the API from fitting only a single input; and a final publishing agent checks whether the input and output structures are reasonable. For login and two-factor authentication, the platform switches to a deterministic login flow to prevent models and recorders from reading customer credentials, and it also supports saving TOTP seeds or automatic filling through a Pomerado phone number.
The team said that when a run fails in the future, it will launch maintenance tasks to repair broken steps such as a renamed button, and may also add functionality or new safeguard rules. Team members said they had previously taken part in agent deployments in healthcare, professional services, and telecommunications at McKinsey and Scale AI, where they found traditional websites and application interfaces to be a common bottleneck.