Finance ops · Production

Ramp builds MCP server enabling natural language querying of business spend data via Claude

The problem

Ramp's initial MCP prototype ran into scaling issues—miscalculations, limited context windows, input size limits, and high token usage—preventing it from handling more than a few hundred transactions when connecting Claude to business spend data.

First attempt

A simple pagination tool built to chunk API responses into smaller parts did not solve the scaling problem.

Workflow diagram · grounded in source
1
Natural language query submitted
trigger
“With Claude Desktop as our MCP client, Claude was able to both generate visualizations and run simple analyses on spend data pulled from Ramp's APIs using natural language”
2
Transactions loaded from Ramp API
integration
“load_transactions) to pull data from the Ramp API”
3
Data transformed into SQLite
integration
“process_data tool to transform the data from the API and load into the SQLite table”
4
SQL queries executed by Claude
ai_action
“execute_query to run queries on the in-memory database directly”
5
Analysis and visualization delivered
output
“Claude was able to both generate visualizations and run simple analyses on spend data pulled from Ramp's APIs using natural language”
Reported outcome

After switching to a SQL-based paradigm with in-memory SQLite, Claude went from struggling with a few hundred data points to accurately analyzing tens of thousands of spend events.

Reported metrics
Scale of spend events analyzedfrom struggling with a few hundred data points to accurately analyzing tens of thousands of spend events
User reaction to resultsmind-blowing
Reported stack
FastMCPClaude DesktopSQLiteClaude
Source
https://engineering.ramp.com/post/ramp-mcp
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After switching to a SQL-based paradigm with in-memory SQLite, Claude went from struggling with a few hundred data points to accurately analyzing tens of thousands of spend events.

What tools did this team use?

FastMCP, Claude Desktop, SQLite, Claude.

What results were reported?

Scale of spend events analyzed: from struggling with a few hundred data points to accurately analyzing tens of thousands of spend events; User reaction to results: mind-blowing (source-reported, not independently verified).

What failed first in this deployment?

A simple pagination tool built to chunk API responses into smaller parts did not solve the scaling problem.

How is this finance ops AI workflow structured?

Natural language query submitted → Transactions loaded from Ramp API → Data transformed into SQLite → SQL queries executed by Claude → Analysis and visualization delivered.

Related finance ops cases