See how much of a model's context window your MCP servers consume — and which server to cut first when they no longer fit.
| Context window | Usable for tools | Tool tokens | Share | Headroom | Verdict |
|---|---|---|---|---|---|
| 128,000 GPT-4o, GPT-4o Mini, DeepSeek Chat +1 | 88,000 | 30,250 | 24% | 57,750 | Severe tokens: Moderate · count: Severe |
| 200,000 o3, o4-mini, Claude 4 Sonnet +3 | 142,000 | 30,250 | 15% | 111,750 | Severe tokens: Moderate · count: Severe |
| 1,000,000 GPT-4.1, GPT-4.1 Mini, GPT-4.1 Nano | 742,000 | 30,250 | 3% | 711,750 | Severe tokens: Healthy · count: Severe |
| 1,048,576 Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.0 Flash +1 | 778,432 | 30,250 | 3% | 748,182 | Severe tokens: Healthy · count: Severe |
| 10,485,760 Llama 4 Scout | 7,856,320 | 30,250 | 0% | 7,826,070 | Severe tokens: Healthy · count: Severe |
Tool token counts are estimates from a character-based cost model, not measured tokenizer output. Real counts vary by provider tokenizer and by how verbose your tool descriptions are — treat the percentages as a sizing guide, not a billing figure.
Attaching MCP servers has a cost you pay on every single request: each server's tool definitions are injected into the context window before your actual question. A tool definition is mostly its description plus its inputSchema, which is why verbose schemas are expensive — the calculator prices the description at roughly 4 characters per token (prose) and the schema at 3.6 (JSON, denser). Your usable budget is the window minus what you reserve for the model's reply and for the conversation itself. But there is a second, independent failure mode: too many tools. Practitioners recommend no more than 10–15 tools exposed at once, because tool-selection accuracy has been measured collapsing from 43% to under 14% on bloated tool sets — and that happens regardless of how few tokens they consume. A set of 30 terse tools can look perfectly healthy by token share and still wreck your agent. So this calculator grades both dimensions separately and reports the worse of the two.
Tool definitions are charged on every request. The two failure modes — window pressure (by tokens) and selection collapse (by tool count) — are graded independently, and the worse verdict wins.
A developer attaches github (35 tools), slack (12) and filesystem (8) — 55 tools totalling roughly 30,250 tokens at the 550-token default.
Why did the cow cross the road?
No signups, no data sold. The core of every tool is free forever — the optional Pro plan adds batch processing, unlimited downloads, white-label exports and an ad-free experience.
☕Support me on Ko-fi— keep tools free100% of proceeds go towards hosting & building more free tools.