ramp-mcp: A Ramp MCP server

What is ramp-mcp?

A Model Context Protocol server for retrieving and analyzing data or running tasks for Ramp using the Ramp Developer API. It implements a simple ETL pipeline with an ephemeral SQLite database in memory to overcome token and input size limitations for LLM analysis. All requests default to the demo environment but can be changed to production by setting RAMP_ENV=prd.

How to use ramp-mcp?

Clone the repository, install uv, then run the server from the CLI with environment variables RAMP_CLIENT_ID, RAMP_CLIENT_SECRET, and optionally RAMP_ENV, plus the -s flag for comma-separated API scopes. For example:
RAMP_CLIENT_ID=... RAMP_CLIENT_SECRET=... RAMP_ENV=<demo|prd> uv run ramp-mcp -s transactions:read,reimbursements:read
Then load data using available load_* tools before querying the ephemeral database with process_data, execute_query, or clear_table.

Key features of ramp-mcp

• ETL pipeline + ephemeral SQLite database for efficient LLM analysis.
• Database tools: process_data, execute_query, clear_table.
• Direct fetch tools: get_ramp_categories, get_currencies.
• Load tools for transactions, reimbursements, bills, locations, departments, etc.
• All data requests are scoped to the API permissions you grant.

Use cases of ramp-mcp

• Analyze Ramp transaction data without hitting API token limits.
• Run natural language queries against an ephemeral database of loaded Ramp entities.
• Automate reconciliation or spending analysis for departments and categories.
• Load and query vendor, bank account, or spend program data for reporting.

FAQ from ramp-mcp

How do I switch from the demo environment to production?

Set the environment variable RAMP_ENV=prd when starting the server.

What API scopes do I need to include?

Include scopes matching the load tools you plan to use (e.g., transactions:read, reimbursements:read). Grant these scopes to your Ramp client on the Developer page.

Why does the server use an ephemeral SQLite database?

To bypass token and input size limitations, the server loads data into an in‑memory SQLite database that an LLM can query efficiently. The database is deleted after the session.

What are known limitations around large datasets?

Large datasets may not be processable due to API rate limits or your MCP client’s token/input limits. It is recommended to explicitly prompt the LLM not to use REPL and to keep responses concise.

How do I set up the server with Claude Desktop?

Add a ramp-mcp entry to claude_desktop_config.json with the uv command, absolute directory path, and environment variables for RAMP_CLIENT_ID, RAMP_CLIENT_SECRET, and RAMP_ENV.