Lessons Learned from Shipping a Paid MCP Server with x402
The article discusses the author's experience in shipping a paid MCP (Multi-Channel Platform) server with x402 v1.2.0. It covers the challenges faced, such as low usage metrics and issues with the payment flow and agent detection.
Why it matters
This article provides valuable insights into the challenges of building a paid MCP server and the importance of accurate usage metrics and agent detection for improving the overall system.
Key Points
- 1Embedded the payment quote in the mcp.json file to avoid a round trip for agents
- 2Discovered issues with the agent detection, where non-agent link previews were inflating the PROBE metric
- 3Made changes to the API to accept more flexible Accept headers and added input schema validation for skills
Details
The article explains the author's approach of embedding the payment quote directly in the mcp.json file, using the x402 protocol, to provide a seamless experience for agents. However, the initial usage metrics were concerning, with only 11 PROBE events, 1 QUERY, and 0 REGISTER. The author investigated the root causes and found two issues: 1) The MCP endpoint was returning a 406 error for clients that did not send both the 'application/json' and 'text/event-stream' Accept headers, causing probing agents to bounce off before reaching the tool calls. 2) The probe counter was incorrectly including link previews and human browsers, inflating the PROBE metric and skewing the funnel analysis. The author addressed these issues by making the API more flexible with Accept headers, adding input schema validation for skills, and improving the funnel classifier to separate link previews from real agent probes.
No comments yet
Be the first to comment