Yes, we run MCPSpend on MCPSpend.

How we cut our own Claude + Cursor bill by 38% in 3 weeks using the tool we built.

“I built MCPSpend because the existing AI observability tools couldn't tell me which MCP tool was eating my budget. Three weeks in, the answer turned out to be 70% Playwright. So I changed how my agents use it and saved $5.46/week — on the very product whose pricing page I was building.”

The challenge

Building MCPSpend itself meant running a half-dozen MCP servers — Playwright for end-to-end tests, GitHub for code search, filesystem + sqlite for state, fetch for API checks — across Cursor, Claude Desktop, Windsurf, and Claude Code simultaneously. The total bill from Anthropic + OpenAI kept climbing past $14/week, but I had no idea where it was going. The very gap I built MCPSpend to fix.

The approach

1. 1Installed the proxy with `npx @mcpspend/proxy add` — three IDEs auto-detected, eight MCP servers wrapped, zero config changes on my side.
2. 2Watched the Live Ticker for a week to baseline normal usage.
3. 3Drilled into the top-tools dashboard view — Playwright `browser_navigate` was 70% of cost.
4. 4Looked at session details for the worst offenders: most browser_navigate calls returned full DOM snapshots that got stuffed into context.

Results

• ✓Refactored my QA agent to scrape only the relevant DOM subtree instead of full page — Playwright cost dropped 60%.
• ✓Added a simple read-file cache around filesystem MCPs that re-read the same files within a session — halved filesystem cost.
• ✓Set a Slack budget alert at $10/week so I notice drift early.
• ✓Net weekly spend down from $14.20 to $8.74 — a 38% cut without changing what the agents do.

Stack