Ramp built an open-source MCP server that lets LLMs interact with business spend data using natural language. Built with Python's FastMCP library, the server exposes Ramp's developer API as MCP tools. Initial prototyping with Claude Desktop revealed scaling issues with large datasets, which were solved by implementing an in-memory ETL pipeline that loads API data into SQLite, allowing Claude to query data via SQL instead of processing raw JSON. This dramatically improved accuracy and scalability from hundreds to tens of thousands of transactions. The post covers the architecture, tooling design, security considerations (OAuth scopes, audit logging, least privilege), and remaining limitations like API latency and reliability of write tools.
Table of contents
What MCP is and how we used itLLMs are surprisingly good at SQLWhere Ramp MCP shines (and where it breaks down)Ramp MCP and Audit LogsConclusions2 Impressions