Claude Code Is Burning Your Token Budget. Here's the Receipt.

I found $2,500/year of hidden token waste in my Claude Code setup. It was the MCP servers.

Last week I noticed my Claude Code conversations were dying at around message 15. Context window full. The model starts forgetting earlier instructions. Tool calls fail. The conversation degrades into hallucination.

I assumed it was my fault — too many messages, too much context. So I started measuring.

Here's what I found:

Session start:
  Claude system prompt:        ~8,000 tokens
  MCP schema injection:      ~111,000 tokens
  User's first message:            50 tokens
  ──────────────────────────────────────────
  Total before any work:     ~119,000 tokens

  Remaining context:         ~81,000 tokens

I was starting every conversation with 60% of my context already consumed.

The culprit wasn't my prompts. It was the 10 MCP servers I had proudly configured in my claude_desktop_config.json

.

I measured each server's schema injection using tiktoken:

Look at the "Times Used/Week" column. Three servers were used zero times. Two more were used once or twice. But every single one of them was injecting 100% of its schema into every conversation.

I was paying $0.33 per conversation — $2,500/year — to load schemas for tools I barely used.

I posted my findings on Bluesky. Within hours:

"I had the same issue. Removed 6 MCP servers and my conversations went from dying at message 15 to lasting 40+ messages." —

[@developer1]"GitHub MCP is 12K tokens but Claude Code already has

gh

CLI built in. Why did I install it?" — @developer2"Google Drive MCP alone is 47K tokens. FORTY SEVEN THOUSAND. For a tool I used once this month." — @developer3

This isn't a niche problem. Everyone running 5+ MCP servers is silently burning 50-100K tokens per conversation.

pip install mcptoon
mcptoon audit --config ~/.config/claude/claude_desktop_config.json

Output:

Filesystem:     3,847 tokens  ✓ Keep (used daily)
GitHub:        12,440 tokens  ✗ Remove (use `gh` CLI instead)
Google Drive:  47,293 tokens  ✗ Remove (use `gdrive` CLI)
Slack:         14,672 tokens  ✗ Remove (use `slack` CLI)
Notion:        13,780 tokens  ⚠ Depends (no good CLI alternative)
Postgres:       8,231 tokens  ✓ Keep (used 4x/week)
...
Total waste:   89,702 tokens  →  $1,346/year
{
  "mcpServers": {
    "filesystem": { ... },
    "postgres": { ... },
    "brave-search": { ... }
  }
}

I went from 10 servers to 3. My context overhead dropped from 111K to 14K tokens.

For servers you must keep, wrap them with mcptoon to compress schemas:

{
  "mcpServers": {
    "postgres": {
      "command": "mcptoon",
      "args": ["serve", "--stdio", "npx", "@modelcontextprotocol/server-postgres"]
    }
  }
}

mcptoon compresses tool schemas by 97% using a compact TOON format. The model sees the same tools but with 3K tokens instead of 111K.

This isn't just about money. The real cost of MCP schema bloat is cognitive degradation:

When Perplexity's CTO announced they were replacing MCP with REST API + CLI internally, this is what he was talking about. When Garry Tan said "MCP sucks honestly," this is what he meant. When Anthropic's own engineers showed 98.7% token reduction by not schemas into context, this is the problem they were solving.

MCP servers aren't bad. The protocol isn't bad. But ** every tool schema into context at startup is a design flaw that costs every developer thousands of dollars per year in wasted tokens.**

Most developers don't know this because:

Now you know. Go audit your config.

pip install mcptoon
mcptoon audit --config your-config.json

mcptoon is open source, Apache 2.0, 250KB, zero dependencies. GitHub · PyPI. All measurements use tiktoken cl100k_base. Not affiliated with Anthropic.