mcpfold logomcpfold

MCP token calculator

The MCP token calculator estimates how many tokens your connected MCP servers spend on tool definitions every turn — and how much per-client curation with mcpfold saves. Every server you connect loads its full tool schema into the model’s context on each turn, whether the agent uses those tools or not. Everything runs in your browser; nothing is uploaded.

Load your own config (optional)

Paste a client’s MCP config (the mcpServers block from Claude, Cursor, VS Code, …). It’s parsed locally to read your server names; set each tool count below.

Your servers

tools
tools
tools
tools
Quick-add:
Tools loaded
5616
Tool-schema tokens
9,3402,676
Share of a 200K window
4.7% → 1.3%
Context cut
71%
6,664 tokens / turn
Stop paying that tax. mcpfold curates the toolset per client from one canonical config — deterministically, on every client, with no extra config.
Install mcpfold

How this is estimated: tokens are computed from a representative tool schema (~1 token ≈ 4 characters of JSON, the common rule of thumb), the same method as mcpfold’s reproducible benchmark, which also publishes exact per-model counts (GPT and Claude tokenizers) that land within a few percent of this approximation. Real cost depends on each tool’s actual schema, so treat the tool counts as editable estimates. The relative reduction is stable regardless of tokenizer. See how to reduce MCP token usage for every approach compared.

Frequently asked

How many tokens do MCP servers use?

Each MCP server loads its full tool schema into the model’s context on every turn, whether the agent uses those tools or not. A single busy server’s tool definitions can run to thousands of tokens, and a handful of servers can consume a large share of the context window before the agent does any work. The exact cost depends on each tool’s JSON schema; the mcpfold token calculator estimates it from a representative schema so you can size your own setup.

How do I reduce the tokens my MCP tools use?

Load only the tools each agent needs. mcpfold curates the toolset per client from one canonical config, so unused tool schemas never enter the context window. In a reproducible benchmark, trimming a 45-tool setup to the 9 tools actually needed cut tool-schema tokens by about 80%. This works on every client, including Cursor, Windsurf, and Zed, which have no native tool-search.

Is the MCP token calculator accurate?

The mcpfold token calculator gives a representative estimate using the common 1-token-≈-4-characters rule on a typical tool schema — the same method as the published benchmark. Real cost varies with each tool’s actual schema, so tool counts are editable; the relative reduction from curation is stable regardless of the tokenizer.