A new browser-based agent automates the creation of AI tools by reverse-engineering authenticated web application APIs. This approach bypasses the need for developers to manually create or maintain API integrations, effectively generating an "MCP server" that adapts to host application changes.

The agent observes how an application interacts with its own APIs, extracting details such as endpoint, method, authentication protocols, and schema definitions. These extracted components are compiled into "recipes," which serve as reusable tools for LLMs. This method allows AI assistants to perform actions directly within applications, like inviting users to a workspace, without relying on brittle browser automation or slow token-burning methods.

This system addresses the common challenge of integrating AI agents with modern software, which often feature complex, poorly documented APIs and inconsistent security standards. The project demonstrates functionality with applications like Jira and Spotify, enabling deeper product integration beyond typical chat or RAG functionalities. The developers noted that standardizing recipes from GraphQL APIs proved particularly challenging.