Designing an MCP server around an existing 600-endpoint API (parallellabs.app)

Parallel AI has announced a significant engineering achievement in optimizing their platform's API design to better accommodate a growing number of AI agents and tools. With over 600 operations in their API, they faced the challenge of ensuring unified access across various callers: their web app, developers using REST APIs, external MCP clients, and in-app AI agents. The team designed a streamlined architecture that allows for a single source of truth in their API routes, minimizing redundancy in authentication and documentation while ensuring consistency across all operational calls. Instead of managing separate implementations and documentation for each caller type, they utilize a Flask route system that dynamically generates OpenAPI specifications, allowing all callers to access the same endpoint with robust authorization. This streamlined approach is particularly impactful for the AI/ML community as it simplifies integration for developers while ensuring accurate, real-time API documentation generated from existing code. By using decorators to attach metadata and building the OpenAPI spec on-the-fly, Parallel AI not only enhances operational efficiency but also maintains a clear and easy-to-navigate catalog of API functionalities. The decision to present just four tools to external MCP clients—versus 603 operations—illustrates a proactive approach to reduce complexity for developers and AI agents while maintaining comprehensive functionality. This innovation has the potential to influence best practices in API design for other companies in the AI/ML space.