MCP Tool Context Overhead Calculator

An MCP tool is not free the moment you call it — it is charged the moment it is available. Tool definitions live in the request prefix, so the full JSON schema of every tool on every connected server is serialised into every API call for the entire session. Twenty servers you never invoke still cost you the same tokens as twenty servers you use constantly.

Per-tool cost

Each tool definition is a name, a natural-language description, and a JSON Schema for its arguments. Its serialised length in characters, divided by the tokenizer's characters-per-token ratio, gives the token cost:

The default k = 3.6 reflects that JSON is denser than prose: braces, quotes, colons and repeated keys such as "type" and "properties" break into many short tokens, so a schema yields more tokens per character than English text, which sits nearer 4 characters per token for Claude's tokenizer. That is why CLAUDE.md is divided by 4 here while schemas are divided by 3.6. Both are estimates — the only exact figure comes from Anthropic's token counting endpoint, or from running /context inside Claude Code, which prints the real live breakdown.

Server and total overhead

F is fixed overhead: it is present before the first user message and it never shrinks during the session. Two numbers follow immediately:

Turns until auto-compact

Claude Code watches how full the transcript is and triggers auto-compact as it approaches the ceiling — the threshold here defaults to 92 percent of the window and is adjustable, because the exact trip point moves between releases. Conversation grows by roughly a constant a tokens per turn (your message, the reply, and any tool results), so compaction happens on the first turn where:

Because F sits in the numerator, every token of tool schema is a token stolen directly from your turn budget. Disabling a server with T_s tokens hands those tokens back, and the turns you regain are approximately T_s / a — the payoff table computes the exact integer difference by re-solving the equation with and without each server.

Compactions over a session

This is the number that actually hurts. Every compaction is a summarisation pass that discards detail, loses file contents you had already loaded, and frequently makes the model repeat work it has already done. Cutting fixed overhead is the cheapest way to reduce it: it costs nothing but a config edit.

Ranking, and what to do about it

• Tokens per tool exposes verbose schemas. A server averaging 800 tokens per tool is spending them on long descriptions and deeply nested argument objects; one averaging 200 is lean. High per-tool cost is the signal to trim, not necessarily to remove.
• Total T_s is what you actually pay. A 90-tool server dwarfs everything else even with modest per-tool cost, which is why the largest integrations ship toolset flags — enabling only the toolsets you need is usually a bigger win than dropping the server.
• Prompt caching makes this prefix cheap in dollars, since a stable tool block is a cache hit on every request. It does not make it cheap in context: cached tokens still occupy the window. Cost and capacity are separate problems, and this calculator is about capacity.
• Practical levers, in order of payoff: turn off servers you are not using this session; enable narrow toolsets instead of all of them; shorten tool descriptions in servers you control; and keep CLAUDE.md tight, since it is re-sent on every request too.

Verify against reality: run /context in Claude Code for the live breakdown, and use each server's own tools/list output in the exact mode above rather than the library estimates.