MCPSpendHow to track Claude Desktop costs (the missing observability layer)
Anthropic shows you total usage; it doesn't show you which MCP tool spent it. Here's how to attribute every dollar to its tool, server, and project.
If you use Claude Desktop with MCP servers (filesystem, GitHub, Playwright, custom), you've probably noticed that Anthropic's usage dashboard tells you how much you spent but not what on. This post walks through how to fix that in 60 seconds.
What Anthropic shows you
Open claude.ai/settings/usage and you'll see:
- Total token usage by month
- Cost per model
- Daily breakdown
What's missing:
- Which MCP tool consumed the tokens
- Which project / conversation / time of day was expensive
- Whether one runaway agent session is dominating
For most users this is fine. For anyone running Playwright MCP, a custom database MCP, or multiple parallel agents, the missing breakdown is exactly where the surprises live.
The minimum setup
Install MCPSpend's proxy:
npx @mcpspend/proxy add
That's it. The CLI auto-detects Claude Desktop's config (claude_desktop_config.json on macOS/Windows) and rewrites it to wrap every MCP server you have. Restart Claude Desktop, make any tool call, and the data lands in your dashboard in real time.
Total time: under a minute. Zero changes to Anthropic's settings.
What you actually get
After 24 hours of normal usage, the dashboard shows you:
- Per-tool cost ranking. Usually 2-3 tools account for 60-80% of spend.
- Per-session drill-down. Click a session → see every tool call it made with cost and latency.
- Budget alerts. Email or Slack when you hit 50/80/100% of a dollar threshold you set.
- Comparison week-over-week. "Playwright cost 3× more than last week" surfaces fast.
A real example
From my own Claude Desktop usage over the last 30 days:
| Tool | Calls | Cost | % of total |
|---|---|---|---|
| playwright/browser_navigate | 1,820 | $3.92 | 40% |
| filesystem/read_file | 984 | $1.71 | 17% |
| github/search_repos | 612 | $1.40 | 14% |
| everything else | 1,996 | $2.81 | 29% |
Three tools = 71% of the bill. Knowing this, I refactored my QA agent to scrape a DOM subtree instead of the full page — Playwright cost dropped 60%.
The free tier
MCPSpend's free plan is 25,000 tool calls per month, no credit card required. For most solo Claude Desktop users that's more than a month of heavy use. If you run automated agents that exceed that, Pro starts at $29/month.
Try it
npx @mcpspend/proxy add
You'll get a dashboard URL within seconds. No SDK changes, no code edits, works alongside Anthropic's existing tooling without interference.
See pricing · Read the security model · Compare with Helicone