Cursor MCP costs: a transparent breakdown across 8 popular MCP servers

Cursor's mcp.json is the simplest way to add MCP servers to any agentic IDE. But once you have 5-10 servers wired up, cost attribution becomes a black box — Cursor itself doesn't tell you which server is the expensive one.

This post breaks down 30 days of real Cursor + MCP usage across the most popular servers I've seen in user installations.

The setup

// ~/.cursor/mcp.json (before MCPSpend wraps it)
{
  "mcpServers": {
    "filesystem":  { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"] },
    "github":      { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"] },
    "playwright":  { "command": "npx", "args": ["-y", "@playwright/mcp"] },
    "sqlite":      { "command": "uvx", "args": ["mcp-server-sqlite", "--db-path", "/path/to.db"] },
    "fetch":       { "command": "uvx", "args": ["mcp-server-fetch"] },
    "postgres":    { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-postgres", "postgres://..."] },
    "memory":      { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] },
    "time":        { "command": "uvx", "args": ["mcp-server-time"] }
  }
}

8 servers. That's a typical heavy-Cursor user.

The 30-day breakdown

Top finding: Playwright is 5.4× the next-biggest server. If you use it, that's where to look first.

The per-tool drill-down

Inside Playwright, the cost distribution is even more skewed:

• browser_navigate — 71% of Playwright spend
• browser_take_screenshot — 14%
• browser_click — 8%
• everything else — 7%

browser_navigate returns the full page DOM by default. For a moderately complex site that's 80-150 KB of HTML → 20K-40K tokens stuffed into context every time. Three optimizations that helped:

1. Scrape DOM subtree, not full page. Use browser_evaluate to grab document.querySelector('main').outerHTML instead — typically 5-10× smaller payload.
2. Cache navigations within a session. If your agent visits the same URL twice in 60 seconds, the second browser_navigate is wasted.
3. Use browser_snapshot for state checks. It returns accessibility tree (text-only), not HTML — way cheaper.

How to do this analysis yourself

npx @mcpspend/proxy add

That's the entire setup. The CLI auto-detects Cursor's config, wraps every MCP server, and starts streaming data. Open the dashboard, wait a week of normal use, and the same table above appears for your usage.

Free tier covers 25,000 tool calls/month — that's a heavy Cursor user.

Counterpoint: when not to do this

If you only use 1-2 MCP servers and your monthly bill from Anthropic is under $20, this is overkill. The value kicks in around 5+ servers or $30+/month in Cursor usage. Below that, the manual mental model works fine.

Try it

npx @mcpspend/proxy add · MIT licensed · pricing · security