Originally published on Serverdex, where the numbers update live.
Here's a cost most people never see: every MCP tool definition is loaded into your model's context before it does any work. Connect a server, and its tools' names, descriptions and full argument schemas are sent to the model on every single turn — before it reads your prompt.
Ten lean tools are cheap. One server with two hundred verbose tools is a standing bill you pay on every message. Nobody had measured that bill across the ecosystem, so I did.
I called tools/list on every public MCP server that answers — no API key, no OAuth — and measured the tool and token footprint. From the live run on 20 Nov 2026 (13,716 answering servers):
- The median MCP server loads ≈4,847 tokens of tool definitions — 14 tools — before anything happens.
- The top 10% cost ≈40,000 tokens.
- The heaviest single server costs ≈240,000 tokens — more than a whole 200K context window, on its own.
Why it hurts three ways
That bill isn't just tokens. It's:
- Money — you pay for those input tokens every turn.
- Latency — more context is slower to process.
- Attention — past a certain size, models start to overlook tools or pick the wrong one. (It's the same thing the top-voted MCP post on Hacker News was about: cutting a server's context footprint.)
The distribution is lopsided
Most answering servers sit in the 6–50 tool range, but the tail is heavy:
| Tools exposed | Servers |
|---|---|
| none listed | 67 |
| 1–5 | 2,890 |
| 6–20 | 4,904 |
| 21–50 | 5,282 |
| 51+ | 573 |
The jump from the median (~4.8k tokens) to the top 10% (~40k) is roughly 8×. A small number of servers carry most of the weight. The longest tool lists I found: minia2a (1,932 tools), Novarithm Agent Hub (1,040), DelegAIze (922).
What it costs your context window
Set against a 200K-token window (Claude's, used here only for scale):
- One median server: ≈4.8k tokens — ~2% of the window.
- Five median servers connected at once: ≈24k tokens — ~12% of the window spent on definitions before you type a word.
- The heaviest single server: wouldn't fit at all.
How to cut your context tax
You rarely need every tool of every server in one task. From biggest effect to smallest:
- Connect fewer servers per task. This is the whole game. A research task doesn't need your database, your deploy tooling and your design server all loaded at once.
- Expose only the tools you need. VS Code's
toolsfield, Cursor's per-server enable switch, Cline's per-server toggles. Turning off the 20 tools you never call on a 30-tool server reclaims most of its bill. - Prefer lean servers. When two servers do the same job, the one with ten focused tools beats the one with eighty.
- Watch the symptom. If a model suddenly gets slow or ignores tools, suspect the context first.
Method & data
Serverdex calls every listed server with a streamable-HTTP remote about once a day with a credential-free initialize + tools/list, and records the tool list. Token cost is an estimate: characters of the raw definitions ÷ 4 (the floor — a real client frames each one with more). Percentiles are over answering servers.
- Full study (updates live): https://serverdex.io/articles/mcp-context-tax
- Raw data (CSV, free to reuse with attribution): https://serverdex.io/state-of-mcp/data.csv
- The broader numbers: The State of MCP Servers
- Servers answering right now: /remote
If you maintain MCP servers: your tool list is a budget. Spend it like one.