After tracking 30 days of MCP tool calls, here's what I learned

I spent the last 30 days logging every MCP tool call I made across
Cursor, Claude Code, and Claude Desktop. Some surprises:

1. Read tool calls dominate by volume but tokens are tiny — they're
   cheap. The cost spike comes from a handful of big calls (web search,
   image processing, large file reads).

2. ~14% of tool calls fail or timeout. They still cost something on
   the LLM side because the model already burned input tokens.

3. There's no native way in Cursor/Claude/Windsurf to see which
   project or which session caused the spend. You only see totals.

4. Per-agent attribution flips assumptions: "the Claude Code in my
   dev folder" vs. "the Cursor in my marketing folder" had a 7x cost
   delta for the same number of files touched.

I turned the logger into a product (MCPSpend) and we're launching on
PH tomorrow. Free tier 25k calls/month if anyone wants to test:
npx --yes @mcpspend/proxy@latest init --key mcps_live_xxx

Curious — anyone else tracking their MCP spend manually? What's
your stack?