Workflow · Production

Pricing the Unknown: Ref builds credit-based pricing for the first standalone paid MCP server

The problem

Ref faced the novel challenge of pricing a first-of-its-kind standalone paid MCP server in an ecosystem where nearly everything is free, while covering high web-crawling and indexing costs and supporting usage volumes ranging from individual developers making ten searches a day to agents making thousands.

Workflow diagram · grounded in source
1
AI agent calls MCP tool
trigger
“Our MCP server has two tools, (ref_search_documentation and ref_read_url ), each billed at 1 credit per call.”
2
LLM retrieval processing
ai_action
“AI processing - LLM costs in our retrieval pipeline, in addition to the base Turbopuffer query costs.”
3
Index search via Turbopuffer
integration
“Turbopuffer hosts our search indices, charging for both data storage and query volume.”
4
Precise documentation result returned
output
“It helps coding agents find the exact information they need without wasting context window.”
5
Web crawler keeps docs current
feedback_loop
“compute resources for running the web crawler that keeps documentation current”
Reported outcome

Three months after launch, Ref has thousands of weekly users and hundreds of subscribers, with the monthly minimum covering fixed indexing costs and additional credit purchases scaling with usage.

Reported metrics
Weekly usersthousands of weekly users
Paid subscribershundreds of subscribers
Reported stack
Turbopufferref_search_documentationref_read_urlCursorClaude Code
Source
https://www.pulsemcp.com/posts/pricing-the-unknown-a-paid-mcp-server
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Three months after launch, Ref has thousands of weekly users and hundreds of subscribers, with the monthly minimum covering fixed indexing costs and additional credit purchases scaling with usage.

What tools did this team use?

Turbopuffer, ref_search_documentation, ref_read_url, Cursor, Claude Code.

What results were reported?

Weekly users: thousands of weekly users; Paid subscribers: hundreds of subscribers (source-reported, not independently verified).

How is this workflow AI workflow structured?

AI agent calls MCP tool → LLM retrieval processing → Index search via Turbopuffer → Precise documentation result returned → Web crawler keeps docs current.