Skip to content

Commit 499bd89

Browse files
mnifzied-createclaude
mnifzied-create
and
claude
committed
Add per-server MCP token cost reference page (AEO)
Dedicated reference: how many tokens each of 11 MCP servers adds per turn (GitHub 3,546 ... AWS KB 101), with FAQPage JSON-LD targeting exact-match 'how many tokens does X MCP server use' queries. Linked from the study. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 parent adb9637 commit 499bd89

2 files changed

Lines changed: 141 additions & 1 deletion

File tree

‎docs/mcp-token-costs/index.html‎

Lines changed: 139 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,139 @@
1+
<!doctype html>
2+
<html lang="en">
3+
<head>
4+
<meta charset="utf-8" />
5+
<meta name="viewport" content="width=device-width, initial-scale=1" />
6+
<title>MCP Server Token Costs: How Many Tokens Each MCP Server Uses Per Turn — AgentLoop</title>
7+
<meta
8+
name="description"
9+
content="A reference for how many tokens popular MCP servers add to your context every turn. GitHub's MCP server is 3,546 tokens/turn; GitLab 1,194; Slack 679; AWS KB 101. Measured tool-schema overhead across real open-source MCP servers, re-sent on every request."
10+
/>
11+
<link rel="canonical" href="https://mnifzied-create.github.io/agentloop/mcp-token-costs/" />
12+
13+
<meta property="og:type" content="article" />
14+
<meta property="og:title" content="MCP Server Token Costs: How Many Tokens Each MCP Server Uses Per Turn" />
15+
<meta property="og:description" content="GitHub's MCP server adds 3,546 tokens of tool schemas every turn; GitLab 1,194; Slack 679; AWS KB 101. A reference for the per-turn token overhead of popular MCP servers." />
16+
<meta property="og:site_name" content="AgentLoop" />
17+
<meta property="og:url" content="https://mnifzied-create.github.io/agentloop/mcp-token-costs/" />
18+
<meta property="og:image" content="https://mnifzied-create.github.io/agentloop/og-token-tax.png" />
19+
<meta property="og:image:width" content="1200" />
20+
<meta property="og:image:height" content="630" />
21+
<meta name="twitter:card" content="summary_large_image" />
22+
<meta name="twitter:title" content="MCP Server Token Costs: How Many Tokens Each MCP Server Uses Per Turn" />
23+
<meta name="twitter:description" content="GitHub's MCP server adds 3,546 tokens every turn; GitLab 1,194; Slack 679; AWS KB 101. Per-turn token overhead of popular MCP servers." />
24+
<meta name="twitter:image" content="https://mnifzied-create.github.io/agentloop/og-token-tax.png" />
25+
26+
<link rel="stylesheet" href="../styles.css" />
27+
28+
<script type="application/ld+json">
29+
{
30+
"@context": "https://schema.org",
31+
"@type": "FAQPage",
32+
"mainEntity": [
33+
{ "@type": "Question", "name": "How many tokens does the GitHub MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The GitHub MCP server's 26 tool schemas total 3,546 tokens, re-sent on every turn — about $12.89 per 1,000 turns on Claude Sonnet ($4.30 on Haiku). Its fattest single tool is create_pull_request_review (~360 tokens). It is the most expensive MCP server measured." } },
34+
{ "@type": "Question", "name": "How many tokens does the GitLab MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The GitLab MCP server's 9 tool schemas total 1,194 tokens per turn — about $5.83 per 1,000 turns on Claude Sonnet ($1.94 on Haiku). Its fattest tool is push_files (~175 tokens)." } },
35+
{ "@type": "Question", "name": "How many tokens does the Git MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Git MCP server's 12 tool schemas total 1,117 tokens per turn — about $5.60 per 1,000 turns on Claude Sonnet ($1.87 on Haiku). Its fattest tool is git_log (~261 tokens)." } },
36+
{ "@type": "Question", "name": "How many tokens does the Sequential Thinking MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Sequential Thinking MCP server is a single tool that costs 827 tokens per turn — about $4.73 per 1,000 turns on Claude Sonnet ($1.58 on Haiku). That one tool is larger than the entire toolset of most multi-tool servers, because of a ~565-token natural-language description." } },
37+
{ "@type": "Question", "name": "How many tokens does the Slack MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Slack MCP server's 8 tool schemas total 679 tokens per turn — about $4.29 per 1,000 turns on Claude Sonnet ($1.43 on Haiku). Its fattest tool is slack_reply_to_thread (~124 tokens)." } },
38+
{ "@type": "Question", "name": "How many tokens does the Google Maps MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Google Maps MCP server's 7 tool schemas total 547 tokens per turn — about $3.89 per 1,000 turns on Claude Sonnet ($1.30 on Haiku). Its fattest tool is maps_distance_matrix (~124 tokens)." } },
39+
{ "@type": "Question", "name": "How many tokens does the Puppeteer MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Puppeteer MCP server's 7 tool schemas total 538 tokens per turn — about $3.86 per 1,000 turns on Claude Sonnet ($1.29 on Haiku). Its fattest tool is puppeteer_screenshot (~142 tokens)." } },
40+
{ "@type": "Question", "name": "How many tokens does the Brave Search MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Brave Search MCP server's 2 tool schemas total 317 tokens per turn — about $3.20 per 1,000 turns on Claude Sonnet ($1.07 on Haiku). Its fattest tool is brave_web_search (~161 tokens)." } },
41+
{ "@type": "Question", "name": "How many tokens does the Time MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Time MCP server's 2 tool schemas total 237 tokens per turn — about $2.96 per 1,000 turns on Claude Sonnet ($0.99 on Haiku). Its fattest tool is convert_time (~159 tokens)." } },
42+
{ "@type": "Question", "name": "How many tokens does the Fetch MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The Fetch MCP server is a single tool that costs 236 tokens per turn — about $2.96 per 1,000 turns on Claude Sonnet ($0.99 on Haiku)." } },
43+
{ "@type": "Question", "name": "How many tokens does the AWS KB Retrieval MCP server use per turn?", "acceptedAnswer": { "@type": "Answer", "text": "The AWS KB Retrieval MCP server is a single tool that costs 101 tokens per turn — about $2.55 per 1,000 turns on Claude Sonnet ($0.85 on Haiku). It was the leanest MCP server measured." } },
44+
{ "@type": "Question", "name": "Do MCP server token costs stack when you attach several servers?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. Each attached MCP server's tool schemas are sent on every turn, so the costs add up. Attaching GitHub (3,546) + GitLab (1,194) + Slack (679) means ~5,419 tokens of schema re-sent on every request before the user says anything — most of it irrelevant to any single task." } }
45+
]
46+
}
47+
</script>
48+
</head>
49+
<body>
50+
<main class="tp">
51+
<section class="tp-hero">
52+
<div class="badge">Reference · measured tool-schema overhead · estimates</div>
53+
<h1>MCP Server Token Costs</h1>
54+
<p class="sub">
55+
Every MCP server you attach sends its <em>full tool schemas</em> on every turn — a fixed token
56+
cost paid on each request, before the model reads the user's question. Here's how many tokens
57+
popular open-source MCP servers add, measured from their actual tool definitions.
58+
</p>
59+
<p class="sub">
60+
From the study <a href="../token-tax/">The Hidden Token Tax</a> (method + caveats there). Costs
61+
are blended-rate estimates over 1,000 turns plus a modeled 150-token reply; token counts use a
62+
GPT BPE tokenizer (o200k_base) as a consistent estimate for Claude.
63+
</p>
64+
</section>
65+
66+
<section class="tp-panel">
67+
<h2>Per-turn token cost by MCP server</h2>
68+
<table class="tp-table">
69+
<thead>
70+
<tr><th>MCP server</th><th>#Tools</th><th>Tokens / turn</th><th>$/1k turns (Sonnet)</th><th>$/1k turns (Haiku)</th></tr>
71+
</thead>
72+
<tbody>
73+
<tr><td>GitHub</td><td>26</td><td>3,546</td><td>$12.89</td><td>$4.30</td></tr>
74+
<tr><td>GitLab</td><td>9</td><td>1,194</td><td>$5.83</td><td>$1.94</td></tr>
75+
<tr><td>Git</td><td>12</td><td>1,117</td><td>$5.60</td><td>$1.87</td></tr>
76+
<tr><td>Sequential Thinking</td><td>1</td><td>827</td><td>$4.73</td><td>$1.58</td></tr>
77+
<tr><td>Slack</td><td>8</td><td>679</td><td>$4.29</td><td>$1.43</td></tr>
78+
<tr><td>Google Maps</td><td>7</td><td>547</td><td>$3.89</td><td>$1.30</td></tr>
79+
<tr><td>Puppeteer</td><td>7</td><td>538</td><td>$3.86</td><td>$1.29</td></tr>
80+
<tr><td>Brave Search</td><td>2</td><td>317</td><td>$3.20</td><td>$1.07</td></tr>
81+
<tr><td>Time</td><td>2</td><td>237</td><td>$2.96</td><td>$0.99</td></tr>
82+
<tr><td>Fetch</td><td>1</td><td>236</td><td>$2.96</td><td>$0.99</td></tr>
83+
<tr><td>AWS KB Retrieval</td><td>1</td><td>101</td><td>$2.55</td><td>$0.85</td></tr>
84+
</tbody>
85+
</table>
86+
<p class="tp-pricing-note">
87+
Schema-only overhead (the fixed input re-sent every turn). Real bills are marginally higher
88+
(provider tool-block scaffolding not modeled). Measure your own stack with the
89+
<a href="../">free Agent Token Profiler</a>.
90+
</p>
91+
</section>
92+
93+
<section class="tp-panel tp-faq">
94+
<h2>How many tokens does each MCP server use?</h2>
95+
<h3>GitHub MCP server</h3>
96+
<p>26 tool schemas total <strong>3,546 tokens</strong> per turn — ~$12.89 / 1k turns on Sonnet ($4.30 Haiku). Fattest tool: <code>create_pull_request_review</code> (~360). The most expensive measured.</p>
97+
<h3>GitLab MCP server</h3>
98+
<p>9 tools, <strong>1,194 tokens</strong> per turn (~$5.83 / 1k Sonnet). Fattest: <code>push_files</code> (~175).</p>
99+
<h3>Git MCP server</h3>
100+
<p>12 tools, <strong>1,117 tokens</strong> per turn (~$5.60 / 1k Sonnet). Fattest: <code>git_log</code> (~261).</p>
101+
<h3>Sequential Thinking MCP server</h3>
102+
<p>A single tool at <strong>827 tokens</strong> per turn (~$4.73 / 1k Sonnet) — larger than the entire toolset of most multi-tool servers, thanks to a ~565-token description.</p>
103+
<h3>Slack MCP server</h3>
104+
<p>8 tools, <strong>679 tokens</strong> per turn (~$4.29 / 1k Sonnet). Fattest: <code>slack_reply_to_thread</code> (~124).</p>
105+
<h3>Google Maps MCP server</h3>
106+
<p>7 tools, <strong>547 tokens</strong> per turn (~$3.89 / 1k Sonnet). Fattest: <code>maps_distance_matrix</code> (~124).</p>
107+
<h3>Puppeteer MCP server</h3>
108+
<p>7 tools, <strong>538 tokens</strong> per turn (~$3.86 / 1k Sonnet). Fattest: <code>puppeteer_screenshot</code> (~142).</p>
109+
<h3>Brave Search MCP server</h3>
110+
<p>2 tools, <strong>317 tokens</strong> per turn (~$3.20 / 1k Sonnet). Fattest: <code>brave_web_search</code> (~161).</p>
111+
<h3>Time MCP server</h3>
112+
<p>2 tools, <strong>237 tokens</strong> per turn (~$2.96 / 1k Sonnet). Fattest: <code>convert_time</code> (~159).</p>
113+
<h3>Fetch MCP server</h3>
114+
<p>A single tool at <strong>236 tokens</strong> per turn (~$2.96 / 1k Sonnet).</p>
115+
<h3>AWS KB Retrieval MCP server</h3>
116+
<p>A single tool at <strong>101 tokens</strong> per turn (~$2.55 / 1k Sonnet) — the leanest measured.</p>
117+
<h3>Do these stack across multiple servers?</h3>
118+
<p>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 <a href="../token-tax/">the full study</a>.</p>
119+
</section>
120+
121+
<section class="tp-cta">
122+
<p>
123+
These are fixed per-turn estimates from each server's tool schemas. To see your <em>own</em>
124+
stack's per-turn cost — across any combination of servers, models, and tools — paste it into the
125+
free profiler.
126+
</p>
127+
<div class="tp-cta-links">
128+
<a class="btn" href="../">★ Free Agent Token Profiler</a>
129+
<a class="btn primary" href="https://ko-fi.com/s/7306cb3140">AgentLoop Pro — token metering + routing, from $9</a>
130+
</div>
131+
</section>
132+
133+
<footer class="tp-footer">
134+
<a href="../token-tax/">← The Hidden Token Tax (full study)</a>
135+
<span>Estimates · o200k_base · verify against your provider's pricing</span>
136+
</footer>
137+
</main>
138+
</body>
139+
</html>

‎docs/token-tax/index.html‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -386,7 +386,8 @@ <h2>The dataset</h2>
386386
<p class="study-sub">
387387
13 real open-source agents, 79 tools. <strong>Overhead / turn</strong> is the fixed payload
388388
re-sent on every request; <strong>$/1k turns</strong> projects that across 1,000 turns plus a
389-
modeled 150-token reply.
389+
modeled 150-token reply. See the <a href="../mcp-token-costs/">per-server MCP token cost reference</a>
390+
for each server's exact numbers.
390391
</p>
391392
<div class="study-table-wrap">
392393
<table class="study-data">

0 commit comments

Comments
 (0)