How much do MCP tools cost in tokens? (and per request)
Updated 2026-10-25. Each number links to its source, with server version and tokenizer where given. API prices from models.json.
When you connect an MCP server to an agent, the name, description and JSON schema of every tool it exposes usually go into the model's context. That text counts as input tokens on every request, even when the agent never calls the tool. Below are measured token counts for popular MCP servers and what one of them costs per 1,000 requests.
MCP tool-schema token counts
| Server | Tools | Tokens | Version / tokenizer | Source |
|---|---|---|---|---|
| Cloudflare API, full schemas | 2,594 | ~1,170,000 (1,170,523) | native MCP | Official: Cloudflare blog, docs, repo |
| Cloudflare API, required params only | 2,594 | ~244,000 (244,047) | native MCP | Official (same sources) |
| Cloudflare Code Mode | 2,594 | ~1,000 | Code Mode | Official (same sources) |
| GitHub MCP | 47 | 59,084 | o200k_base | Third-party: loadline (run 2026-08-18) |
| GitHub MCP | 47 | 86,843 | Claude tokenizer | Third-party: loadline (run 2026-08-18) |
| Playwright MCP | 23 | 3,130 | @playwright/mcp 0.0.75, o200k | Third-party: Port of Context |
| Notion MCP | - | 15,462 | 2.2.1 | Third-party: Port of Context |
| Sentry MCP | - | 13,823 | 0.36.0 | Third-party: Port of Context |
| 20 MCP servers combined | - | 73,323 (36.7% of a 200K window) | CI-measured, o200k | Third-party: tooltax |
| Linear MCP | 42 | ~12,807 (estimate) | ~4 chars/token rule, not a tokenizer | Quandri engineering blog |
| Slack MCP | 12 | ~3,792 (estimate) | ~4 chars/token rule, not a tokenizer | Quandri engineering blog |
tooltax's own CI counts for single servers (o200k): Playwright 24 tools / 3,423 tokens, Notion 17,430, Firecrawl 16,818. These differ from Port of Context because of server version. Loadline found Claude tokenizer counts run 47-70.5% higher than o200k_base, and that GitHub plus Linear takes 57.9% of a 200K window. Linear and Slack figures are character-based estimates, not tokenizer counts.
Because tool schemas sit in the same prefix on every call, they are a good fit for caching; see does prompt caching save money for AI agents?
What the GitHub MCP schema costs per 1,000 requests
If an agent sends the GitHub MCP tool schemas (59,084 tokens, o200k_base) once per request, the schemas alone cost: 59,084 × 1,000 × price per 1M tokens / 1,000,000. Shown at the uncached input price and at the cached-input price (every schema read a cache hit).
| Model | Input $ per 1M | Uncached, 1,000 requests | Cached input $ per 1M | Cached, 1,000 requests |
|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $236.34 | $0.2 | $11.82 |
| Claude Sonnet 5.5 | $2 | $118.17 | $0.1 | $5.91 |
| GPT-6.1 Sol | $2 | $118.17 | $0.1 | $5.91 |
| Claude Haiku 5.5 | $0.1 | $5.91 | $0.01 | $0.591 |
| DeepSeek Flash (V4.1 Flash 1017) | $0.3 | $17.73 | $0.006 | $0.355 |
Schema-only illustration. It leaves out the prompt, tool results, history, output tokens and cache-write fees. The same 59,084-token count is applied to every model for comparison; Claude models count the same schemas as 86,843 tokens with their own tokenizer (loadline), so their real cost would be higher. Prices are standard tier, short context, from models.json (last updated 2026-10-25).
How to cut MCP token costs
- Use prompt caching. Tool schemas rarely change between requests, so most reads can be cache hits at the cached-input price shown above. Model your hit rate in the prompt caching calculator.
- Trim tools. Only enable the servers and tools a task needs. Cloudflare's numbers show the effect: full schemas ~1,170,000 tokens, required params only ~244,000, Code Mode ~1,000.
- Count with your model's tokenizer. The same schemas can be 47-70.5% more tokens on Claude than on o200k_base.
- Add the overhead to your own estimate in the TokenTab calculator, or see how harness preambles change coding agent costs.
FAQ
How many tokens does an MCP server add to each request?
It depends on the server, its version and the tokenizer. Third-party measurements put GitHub MCP (47 tools) at 59,084 tokens with o200k_base or 86,843 with the Claude tokenizer, Playwright MCP (23 tools) at 3,130, Notion at 15,462 and Sentry at 13,823.
Why do token counts for the same MCP server differ between sources?
Server version and tokenizer change the count. The loadline measurement of GitHub MCP found Claude tokenizer counts 47-70.5% higher than o200k_base counts, so always check which version and tokenizer a number uses.
How much can you cut MCP tool token overhead?
Cloudflare reports that across 2,594 API tools, full schemas are about 1,170,000 tokens, required parameters only about 244,000, and its Code Mode about 1,000 tokens, 99.9% less.
Does prompt caching help with MCP tool schemas?
Tool schemas are usually the same on every request, so they fit prompt caching well, and cache hits are billed at the lower cached-input price. Trimming unused tools cuts the token count itself. Use the TokenTab prompt caching calculator to model your hit rate.
Run your own numbers
Open the TokenTab calculator Prompt caching calculator
Related: Coding agent harness cost · How to estimate AI agent costs.