MCP overhead estimator
Each MCP server can add tool definitions and system instructions to the context window. This estimator multiplies assumed overhead per server by count — replace assumptions with measurements from your /context exports.
For engineers running Claude Code or other MCP-capable agents with more than a couple of servers enabled. The math: servers × assumed tokens per server = standing overhead, which is prepended to every request's context before you type anything.
Heuristics current as of .
Estimated always-on overhead: 4800 tokens before your repo prompt.
MCP overhead is only one part of the always-on context bill — check your CLAUDE.md next.
How to read the results
The overhead is a recurring cost, not a one-time one — it rides along on every request in the session. Worked example: at $3/M input, 10,000 tokens of tool definitions costs about $0.03 per request; across a few hundred requests a day it compounds into real money, and it also crowds out room for actual code in the context window. Measure your real number from a /context export rather than trusting the default assumption.
FAQ
- How is the estimate computed?
- Server count × an assumed per-server token figure. Real overhead varies wildly — a server exposing 3 terse tools costs a fraction of one exposing 40 tools with verbose JSON schemas — so the assumption is a placeholder until you measure.
- What are the limitations?
- Some harnesses defer tool loading — schemas are only pulled into context when needed — which this flat multiplication can't model. It also ignores prompt caching, which can make repeated overhead cheaper on providers that discount cache reads (Anthropic prices them at roughly 10% of input per their pricing page).
- How do I reduce MCP overhead?
- Enable only the servers relevant to the task instead of a global always-on loadout, prefer CLI or direct API calls where a tool adds no real capability, and audit which servers you actually invoked in the last week — unused ones are pure overhead.
- When should I use this vs the full audit?
- Use this for a quick order-of-magnitude check on one variable. The full stack audit measures MCP overhead alongside your CLAUDE.md and caching posture, so you can see which of the three dominates your per-request cost.
For your full setup, run the free stack audit — or see the 100-configs report.