The Real Cost of MCP Tool Calls in Claude Code

MCP turns integrations into repeatable agent actions. That repeatability is exactly what makes costs predictable — and painful.

Where the tokens go

1. Tool definitions in the system context (large if you allow dozens of tools).
2. Invocation payloads — especially when tools return unstructured blobs.
3. Re-prompting — the model re-reads prior tool output as input on the next turn.

Quick numbers

Assume ~250 tokens per tool definition — a reasonable midpoint for common servers; measure your own with the MCP estimator. Schemas are re-read on every turn:

Tools registered	Schema tokens per turn	Per 50-turn session	Cost per session @ Claude Opus 4.8 ($5/M input)
5	1,250	62,500	$0.31
10	2,500	125,000	$0.63
20	5,000	250,000	$1.25
30	7,500	375,000	$1.88

Prompt caching softens this — stable schemas can bill at cache-read rates (10% of input) — but only if nothing above them in the prompt busts the prefix. And these figures cover definitions only; payloads and re-prompted tool output stack on top.

Rule of thumb

If you average more than ~10 tool calls per session on an expensive model, you should budget MCP as a first-class line item, not noise.

TokenCheat’s MCP estimator and audit heuristics help teams surface this before finance does.