MCP Server Token Costs
Every MCP server you attach sends its full tool schemas on every turn — a fixed token cost paid on each request, before the model reads the user's question. Here's how many tokens popular open-source MCP servers add, measured from their actual tool definitions.
From the study The Hidden Token Tax (method + caveats there). Costs are blended-rate estimates over 1,000 turns plus a modeled 150-token reply; token counts use a GPT BPE tokenizer (o200k_base) as a consistent estimate for Claude.
Per-turn token cost by MCP server
| MCP server | #Tools | Tokens / turn | $/1k turns (Sonnet) | $/1k turns (Haiku) |
|---|---|---|---|---|
| GitHub | 26 | 3,546 | $12.89 | $4.30 |
| GitLab | 9 | 1,194 | $5.83 | $1.94 |
| Git | 12 | 1,117 | $5.60 | $1.87 |
| Sequential Thinking | 1 | 827 | $4.73 | $1.58 |
| Slack | 8 | 679 | $4.29 | $1.43 |
| Google Maps | 7 | 547 | $3.89 | $1.30 |
| Puppeteer | 7 | 538 | $3.86 | $1.29 |
| Brave Search | 2 | 317 | $3.20 | $1.07 |
| Time | 2 | 237 | $2.96 | $0.99 |
| Fetch | 1 | 236 | $2.96 | $0.99 |
| AWS KB Retrieval | 1 | 101 | $2.55 | $0.85 |
Schema-only overhead (the fixed input re-sent every turn). Real bills are marginally higher (provider tool-block scaffolding not modeled). Measure your own stack with the free Agent Token Profiler.
How many tokens does each MCP server use?
GitHub MCP server
26 tool schemas total 3,546 tokens per turn — ~$12.89 / 1k turns on Sonnet ($4.30 Haiku). Fattest tool: create_pull_request_review (~360). The most expensive measured.
GitLab MCP server
9 tools, 1,194 tokens per turn (~$5.83 / 1k Sonnet). Fattest: push_files (~175).
Git MCP server
12 tools, 1,117 tokens per turn (~$5.60 / 1k Sonnet). Fattest: git_log (~261).
Sequential Thinking MCP server
A single tool at 827 tokens per turn (~$4.73 / 1k Sonnet) — larger than the entire toolset of most multi-tool servers, thanks to a ~565-token description.
Slack MCP server
8 tools, 679 tokens per turn (~$4.29 / 1k Sonnet). Fattest: slack_reply_to_thread (~124).
Google Maps MCP server
7 tools, 547 tokens per turn (~$3.89 / 1k Sonnet). Fattest: maps_distance_matrix (~124).
Puppeteer MCP server
7 tools, 538 tokens per turn (~$3.86 / 1k Sonnet). Fattest: puppeteer_screenshot (~142).
Brave Search MCP server
2 tools, 317 tokens per turn (~$3.20 / 1k Sonnet). Fattest: brave_web_search (~161).
Time MCP server
2 tools, 237 tokens per turn (~$2.96 / 1k Sonnet). Fattest: convert_time (~159).
Fetch MCP server
A single tool at 236 tokens per turn (~$2.96 / 1k Sonnet).
AWS KB Retrieval MCP server
A single tool at 101 tokens per turn (~$2.55 / 1k Sonnet) — the leanest measured.
Do these stack across multiple servers?
Yes — every attached server's schemas are sent every turn. GitHub + GitLab + Slack = ~5,419 tokens re-sent on every request, most of it irrelevant to any single task. The fix: scope toolsets, strip serialization cruft, and load tools on demand. See the full study.
These are fixed per-turn estimates from each server's tool schemas. To see your own stack's per-turn cost — across any combination of servers, models, and tools — paste it into the free profiler.
AgentLoop Pro turns that measurement into a fix — token metering per run and per tool, plus model-routing that sends cheap turns to a cheaper model — alongside the other production patterns (parallel tools, retries, persistent memory, approval gates) in minimal, readable code.