Garry Tan Was Right: "MCP Sucks Honestly." I Have the Token Receipts.
"MCP sucks honestly. Context window eats too much, auth is a mess. I wrote a CLI wrapper in 30 minutes and it works better."
When YC's CEO says this on X, people listen. But nobody had the data to back it up.
Until now.
What Garry Tan, Perplexity's CTO, and 97 Million Downloads Can't Hide
Three things happened in the last 6 months that changed how I think about MCP:
- Peter Steinberger (OpenClaw founder) tweeted: "mcp were a mistake. bash is better."
- Eric Holmes wrote "MCP is dead. Long live the CLI" — it hit HN frontpage
- Denis Yarats (Perplexity CTO) publicly announced they're replacing MCP with REST API + CLI internally
- Garry Tan (YC CEO) replied: "MCP sucks honestly"
The community split into two camps:
- "MCP is dead" — CLI is simpler, cheaper, faster
- "MCP is fine" — 97M downloads, 17K servers, it's the standard
Both are wrong. The problem isn't MCP. The problem is what MCP does to your context window.
The 47,000-Token Problem Nobody Measured
I connected 10 MCP servers to a token counter. Here's what I found:
| MCP Server | Tools | Token Cost | Equivalent |
|---|---|---|---|
| Sequential Thinking | 3 | 890 | This blog post |
| Brave Search | 8 | 2,103 | A short email |
| Filesystem | 11 | 3,847 | A README |
| Memory | 9 | 2,567 | A meeting note |
| Puppeteer | 15 | 5,890 | A chapter of a book |
| Postgres | 19 | 8,231 | A whitepaper |
| GitHub | 28 | 12,440 | A court filing |
| Notion | 24 | 13,780 | A legal contract |
| Slack | 22 | 14,672 | A novella chapter |
| Google Drive | 31 | 47,293 | Half of a novel |
| Total | 170 | 111,713 | A short book |
One MCP server — Google Drive — injects 47,293 tokens into your context before you ask a single question.
The entire works of Shakespeare is 900K tokens. Google Drive's schema is 5% of Shakespeare. For listing files.
The Cost Breakdown (So You Can Get Angry Too)
At Claude 3.5 Sonnet pricing ($3/M input tokens, $15/M output):
| Scenario | Tokens | Cost | Annual Cost |
|---|---|---|---|
| 1 server (minimal) | 3,847 | $0.01/conv | $4.40/yr |
| 3 servers (common) | 14,528 | $0.04/conv | $19.40/yr |
| 5 servers (typical) | 33,061 | $0.10/conv | $44/yr |
| 10 servers (max) | 111,713 | $0.34/conv | $1,496/yr |
| 10 servers + 20 tool calls | ~180,000 | $0.54/conv | $2,376/yr |
Assumptions: 20 conversations/day, 220 working days/year.
$2,376/year. In JSON braces. For tools you might not even use.
What Anthropic's Own Engineers Admitted
Anthropic published a blog showing that complex MCP workflows can consume 150,000+ tokens. Their solution? A 98.7% token reduction by moving tool definitions from context to runtime code.
Translation: The fix for MCP's token waste is to stop using MCP the way MCP was designed.
Cloudflare went further. Their 2,500-endpoint API originally required 1.17M tokens of schema. They compressed it to 1K by exposing just two functions: search and execute.
1,170,000 → 1,000.
That's not an optimization. That's an architecture admission.
The Real Cost Nobody Talks About: Cognitive Degradation
Money is measurable. But the worse cost is what MCP bloat does to your agent's intelligence:
Without MCP bloat (3K tokens overhead):
Context used by schemas: 2%
Context for reasoning: 98%
Conversation longevity: 40+ messages
Tool selection accuracy: high (few tools to choose from)
With MCP bloat (111K tokens overhead):
Context used by schemas: 56%
Context for reasoning: 44%
Conversation longevity: 15 messages
Tool selection accuracy: low (170 tools, model gets confused)
Your agent isn't getting dumber. Your context window is getting fuller.
"But I Need Those MCP Servers!"
Do you? I audited my own setup:
| Server | Token Cost | Used/Week | CLI Alternative |
|---|---|---|---|
| GitHub | 12,440 | 3 |
gh (built into Claude Code) |
| Slack | 14,672 | 0 | slack-cli |
| Google Drive | 47,293 | 1 | gdrive |
| Notion | 13,780 | 2 | API + curl |
| Puppeteer | 5,890 | 0 |
playwright CLI |
Three servers were used zero times per week. Two more had direct CLI equivalents. I was paying $1,346/year for tools I never touched.
After the audit, I kept 3 servers, removed 7, and proxied the remaining 3 through mcptoon:
pip install mcptoon
| Metric | Before | After |
|---|---|---|
| Servers | 10 | 3 |
| Schema tokens | 111,713 | 3,247 |
| Reduction | — | 97% |
| Annual cost | $2,376 | $69 |
| Savings | — | $2,307 |
mcptoon wraps your MCP servers and compresses their schemas into a TOON format — 97% smaller, same functionality, zero dependencies.
The Verdict
Garry Tan was right. Not because MCP is fundamentally broken — it's not. The protocol is fine. But the implementation pattern — loading every tool schema into context at startup — is a tax on every developer using it.
The evidence is overwhelming:
- YC CEO says it sucks
- Perplexity CTO replaced it internally
- Anthropic's own engineers showed 98.7% waste
- 97M downloads but developers are removing servers after 3 months
- My data: 111K tokens of overhead per conversation, $2,376/year wasted
The fix isn't a new protocol. It's a smarter proxy layer.
pip install mcptoon
Or just remove your unused MCP servers. That's free and takes 2 minutes.
"MCP was a mistake" is too strong. But "MCP as currently implemented is burning your token budget" is exactly right.
All data measured with tiktoken (cl100k_base). Scripts: GitHub. Independent, not affiliated with Anthropic, YC, or Perplexity.
Top comments (0)