MCPSpendWhat does a real MCP tool call actually cost?
A teardown of 30 days of Cursor + Claude Desktop usage, broken down per server, per tool, per project. The cost curve is steeper than you think.
Most "AI cost" articles talk in the abstract — "tokens are getting cheaper", "agents are expensive". Useful, but you can't act on it. So I pulled the raw data from my own MCPSpend account for the past 30 days and broke it down by every dimension MCPSpend tracks.
If you run agents in Cursor or Claude Desktop, this is what your numbers probably look like too.
The headline
In 30 days I made 5,412 MCP tool calls across 9 servers in 3 projects. Total estimated cost: $9.84.
That feels cheap. It is, for a solo dev. But it also hides the distribution that matters.
Top tools by cost
| Tool | Calls | Cost | % of total |
|---|---|---|---|
| playwright/browser_navigate | 1,820 | $3.92 | 39.8% |
| filesystem/read_file | 984 | $1.71 | 17.4% |
| github/search_repos | 612 | $1.40 | 14.2% |
| fetch/fetch | 511 | $1.05 | 10.7% |
| sqlite/query | 287 | $0.66 | 6.7% |
| (everything else, 5 tools) | 1,198 | $1.10 | 11.2% |
Three tools account for 71% of the spend. That's a fat-tail distribution most cost tracking is blind to — Helicone or Langfuse would just show you "$9.84 on Claude this month" and call it done.
What this means in practice
When you can see the breakdown, you make different choices:
- playwright/browser_navigate at 39.8% is the biggest signal. Browser automation is verbose — every
navigatecall comes back with a giant DOM snapshot which gets stuffed into context. Could I prune the DOM before returning? Could I cache navigations across sessions? - filesystem/read_file at 17.4% is mostly redundant. Cursor and Claude Desktop both re-read files multiple times in the same session. A small cache on the proxy side could halve this.
- github/search_repos at 14.2% is mostly me asking "where is this in the codebase" — could be answered by a local grep most of the time.
These are 3 optimisations I never would have considered without per-tool cost attribution.
The per-project view
I split usage across 3 projects: work-client-a (browser-heavy QA agent), mcpspend-itself (this dashboard's own dev work), side-experiments.
- work-client-a: $5.84 (59% of spend, but billable so net positive)
- mcpspend-itself: $3.10 (31% — building the product is the product's biggest customer 😅)
- side-experiments: $0.90 (10%)
Per-project breakdown is the killer feature for anyone who:
- Bills clients for agentic work (need to show what each cost)
- Runs multiple internal experiments and wants to know which deserves more budget
- Has a budget split between teams
The cost forecast
MCPSpend projects end-of-month based on linear extrapolation of the daily average. At day 24, I'm projected to hit $12.30 for the month, which is comfortably inside my $25 personal budget.
But three weeks ago I was projected for $30 (because I'd been doing more Playwright work). The Slack alert at 80% gave me a heads-up; I refactored the worst offender; the curve flattened.
That's the value of real-time observability vs end-of-month surprise.
How to do this teardown yourself
If you have MCPSpend installed, the dashboard at mcpspend.com/dashboard shows all of this out of the box — KPI cards, top tools, per-project totals.
If you don't yet, the free tier (25K calls/month, no card) is more than enough for a single-dev profile:
npx @mcpspend/proxy add
It auto-detects every MCP client on your machine, asks for confirmation, and starts logging. Data appears in under 60 seconds after the first tool call.
Next time
Next post (Wed): "How I built the auto-detection logic that handles 5 different MCP-client config formats."