Serverdex Data Report
The MCP context tax
Every MCP tool definition is loaded into your model's context before it does any work. We measured the tool and token cost of every MCP server that answers. The median server spends about 4.8k tokens on definitions before your first prompt; the heaviest, about 241k — more than a whole 200K context window. Here is the real distribution, and how to keep an agent lean.
Serverdex, “The MCP context tax”: median answering MCP server ≈4,847 tokens of tool definitions (Nov 20, 2026, 17:30 UTC). https://serverdex.io/articles/mcp-context-tax- 1. What the context tax is
- 2. The median server, and the heavy tail
- 3. How many tools servers expose
- 4. What it costs your context window
- 5. The token monsters
- 6. How to cut your context tax
- 7. Methodology & data
1. What the context tax is
A tool is not free to have available. It has to be described to the model first.
When a client connects to an MCP server, it calls tools/list and receives, for every tool, a name, a description and a full JSON schema of its arguments. All of that is placed in the model's context on every single turn, before the model reads your prompt or does any work. Ten lean tools are cheap. One server with two hundred verbose tools is a standing bill you pay on every message.
That bill has three costs: money (you are charged for those input tokens each turn), latency (more context is slower to process), and attention — past a certain size the model starts to overlook tools or pick the wrong one. It is the same problem the top-voted MCP post on Hacker News was about: cutting a server's context footprint. Nobody had measured it across the ecosystem. So we did.
2. The median server, and the heavy tail
Across the 13,716 MCP servers answering our live check, the median exposes 14 tools for about 4.8k tokens. The distribution is not flat, though — it has a long, heavy tail.
The jump from the median to the top 10% is roughly 8×: a small number of servers carry most of the weight. 306,864 tools are indexed across the directory in total.
3. How many tools servers expose
Most answering servers sit in the 6–50 tool range. The ends of the curve are where the trouble is: a few expose none, a few expose hundreds.
Among the 13,716 servers that answered. "None listed" answered the handshake but returned an empty tool list.
4. What it costs your context window
Set against a 200K-token context window (Claude's size, used here only for scale), the footprint of the definitions becomes concrete — before you add a system prompt, your files, or a single message.
Share of a 200,000-token window, for scale. Five median servers already spend ≈24k tokens (12%) on definitions before the model reads anything; the heaviest single server, on its own, would not fit.
5. The token monsters
The servers with the longest tool lists. Powerful, but not something you leave connected while doing something else.
Every server's own tool list and estimated context cost is on its page; the ones answering right now are on /remote.
6. How to cut your context tax
You rarely need every tool of every server in one task. Four moves, from biggest effect to smallest:
1. Connect fewer servers per task. This is the whole game. A research task does not need your database, your deploy tooling and your design server all loaded at once. Keep a small set connected and switch per project.
2. Expose only the tools you need. Several clients let you turn individual tools off: VS Code's tools field, Cursor's per-server enable switch, Cline's per-server toggles. Turning off the twenty tools you never call on a thirty-tool server reclaims most of its bill.
3. Prefer lean servers. When two servers do the same job, the one with ten focused tools beats the one with eighty. Each server's page on Serverdex shows its tool count and estimated token cost, so you can compare before you install.
4. Watch the symptom. If a model suddenly gets slow or starts ignoring tools, suspect the context first — it is cause eleven in our troubleshooting guide. The broader numbers behind this report are in The State of MCP Servers, section 8.
7. Methodology & data
Serverdex calls every listed server that advertises a streamable-HTTP remote endpoint with a plain MCP initialize followed by tools/list — no API key, no OAuth, no account — about once a day, and records the tool list that comes back. The figures here cover the servers that answered. They are as of Nov 20, 2026, 17:30 UTC.
Token cost is an estimate: the characters of the raw tool definitions (names, descriptions and argument schemas) divided by four, the rough bytes-per-token rule. It is the floor — a real client frames each definition with extra formatting, so the true cost is a little higher. Percentiles are over answering servers; "median" is p50, "top 10%" is p90. The 200K window is Claude's, used only as a scale; your client's window may differ.
Serverdex, “The MCP context tax” (November 20, 2026). https://serverdex.io/articles/mcp-context-taxFound an error, or want the per-server breakdown? Get in touch.