Skip to content
mcp/skillhub

Data study · updated 2026-10-19

We measured the context tax of 3,204 MCP servers

Every MCP server you connect sends its tool definitions to the model on every turn. We connected to 7,477 published endpoints, listed their tools and counted the tokens.

Servers measured

3,204

Median tokens / turn

1,554

90th percentile

23,285

Tool definitions tokenized

57,752

Findings

  • The median server costs ~1,554 tokens per turn before you type anything. On Claude Sonnet 5.5 that is $3.11 per 1,000 turns uncached, or $0.311 with prompt caching (prices as of 2026-10-06).
  • The tail is long. 10% of servers add more than 23,285 tokens, 1% more than 35,473, and the heaviest adds 567,902 — on every turn.
  • The largest single tool definition is 16,428 tokens: gen_vidsheet_action on GEN.
  • Most hosted servers can't be inspected anonymously. Of 7,477 endpoints listed in public registries, 24% require sign-in or an API key just to list tools, 24% were unreachable, and 44% answered.
Servers by tokens per turn
  • Lean under 1k774 (24%)
  • Moderate 1k–5k1,618 (50%)
  • Heavy 5k–15k424 (13%)
  • Bloated 15k+388 (12%)

Heaviest servers

  1. 1Crawlora567,902 tok · 3642 tools
  2. 2SAP179,145 tok · 453 tools
  3. 3Minia2a105,453 tok · 1711 tools
  4. 4Gapup MCP Public103,784 tok · 279 tools
  5. 5LiveDataLink81,411 tok · 381 tools
  6. 6GEN73,759 tok · 19 tools
  7. 7MaCalculatriceEnLigne71,770 tok · 446 tools
  8. 8DFX Real Estate Intelligence68,685 tok · 134 tools
  9. 9Rokha67,949 tok · 291 tools
  10. 10Asterwise65,571 tok · 118 tools
Full ranking

Leanest servers (3+ tools)

  1. 1Citation Verifier API142 tok · 3 tools
  2. 2Opencrater154 tok · 3 tools
  3. 3SanctionsAI161 tok · 3 tools
  4. 4Trades Runtime161 tok · 4 tools
  5. 5AI Provider Status API171 tok · 3 tools
  6. 6EntityReach Corporate Group Intelligence184 tok · 3 tools
  7. 7Rust Faf189 tok · 3 tools
  8. 8Blog API189 tok · 3 tools
  9. 9SEAR190 tok · 3 tools
  10. 10AgentsKit Registry192 tok · 4 tools
Full ranking

Method

For every server in the official MCP registry and public lists that publishes a hosted endpoint, we performed the MCP handshake over Streamable HTTP (falling back to SSE) without credentials, called tools/list, and tokenized the definitions as clients send them (name, description, input_schema) plus server instructions, using the o200k tokenizer. This is the worst case: clients that defer tool loading pay less. Local (stdio) servers are being added through a sandboxed pipeline. Full details in the methodology.

Use the data

The dataset is free to use under CC BY 4.0 — please link to this page. Each server also has a README badge showing its live token count.

Download CSV (3,204 rows)