How much context do your MCP tools take?
Every connected MCP server sends its tool definitions with each model request, before your prompt. Pick real, measured servers to see the tokens and what they cost.
Tool snapshot 2026-10-13 · Prices checked 2026-10-12 · USD planning estimates.
17,368 tokens of tool definitions on every request
93 tools from 4 servers. At 1,000 requests a day, the model reads 521.04M definition tokens a month before it reads your prompt.
Context window used
Up to 1.7%
Of each model's input limit, on every request. Caching does not free it.
Per 1,000 requests
$1.74–$173.68
Standard input price across the models below.
Per month
$52–$5.2k
Tool definitions alone, before prompts, history, tool results and output.
Cost by model
| Model | Input / 1M tokens | Context used | Per 1,000 requests | Per month |
|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | 1.74% | $1.74 | $52.10 |
| Gemini 3.8 Flash | $0.75 | 1.66% | $13.03 | $390.78 |
| Claude Sonnet 5.5 | $2.00 | 1.74% | $34.74 | $1,042.08 |
| Claude Opus 5.5 | $4.00 | 1.74% | $69.47 | $2,084.16 |
| GPT-6 Astra | $10.00 | 1.65% | $173.68 | $5,210.40 |
Where the tokens come from
-
GitHub (default toolset)10,989 tokens · 45 tools · 244 per tool
-
Playwright3,745 tokens · 25 tools · 150 per tool
-
Filesystem1,681 tokens · 14 tools · 120 per tool
-
Memory953 tokens · 9 tools · 106 per tool
How to reduce it
- Enable only the tools a task needs. GitHub supports toolsets and individual tools. Tool sizes vary, so count the selected definitions.
- Load definitions on demand. Claude Code tool search and provider tool search defer definitions until the model needs them, at the cost of extra search steps.
- Cache the stable prefix. Prompt caching bills repeated definitions at a fraction of the input price, but they still fill the context window.
- Prefer CLI skills where they fit. Agent Skills progressive disclosure keeps only short metadata in context and loads instructions when used.
Method and limitations
- Tokens per request add the measured tool definitions of the selected servers: name, description and input schema, serialized as compact JSON and counted with tiktoken 0.14.0 (o200k_base). Monthly tokens multiply by requests per day and 30 days.
- Cost multiplies tokens by each model's standard input price. Context used divides tokens by the model's input limit. Prices come from the same dated snapshot as the LLM hosting comparison.
- Excluded: caching discounts, instructions, history, tool results, output and reasoning, client wrappers and output schemas. Each provider tokenizes differently; measure your own client before budgeting.
Sources
- GitHub (default toolset) · v1.12.2
45 tools · 10,989 tokens · Measured 2026-09-20. Default catalog from the pinned public benchmark; not every available GitHub tool.
Pinned measurement source - Filesystem · 1.0.0
14 tools · 1,681 tokens · Measured 2026-10-13. Official reference server; tools/list only, with an empty allowed directory.
Package release - Memory · 1.0.0
9 tools · 953 tokens · Measured 2026-10-13. Official reference server; tools/list only, with a scratch memory path.
Package release - Playwright · 0.0.83
25 tools · 3,745 tokens · Measured 2026-10-13. Default tools/list; no browser launched. Package embeds Playwright 1.64.0-alpha-1790635538000.
Package release
Input prices checked 2026-10-12: Claude Haiku 5.5, Gemini 3.8 Flash, Claude Sonnet 5.5, Claude Opus 5.5, GPT-6 Astra. Pinned public MCP context benchmark.