Les descriptions d'outils MCP sont des coûts cachés. Chaque outil que votre agent peut appeler ajoute une description à l'invite système, consommant des jetons de contexte avant même que l'utilisateur ne dise quoi que ce soit. Avec plusieurs serveurs MCP, les descriptions d'outils peuvent occuper 10 à 20 % de votre fenêtre de contexte. Ce tutoriel construit un outil d'audit de jetons qui mesure exactement combien de jetons chaque serveur MCP consomme, identifie les plus gros consommateurs et montre comment passer à Scavio (mcp.scavio.dev/mcp) consolide 6 plateformes en un seul serveur.
Prérequis
- Python 3.9+ installé
- bibliothèque tiktoken installée (pip install tiktoken)
- Accès à vos fichiers de configuration MCP
- Une clé API Scavio de scavio.dev pour comparaison
Parcours
Étape 1: Charger et analyser votre configuration MCP
Lisez vos fichiers de configuration MCP et listez tous les serveurs configurés et leurs nombres d'outils.
Python
import json, os
def load_mcp_configs() -> dict:
"""Load MCP configs from all standard locations."""
configs = {}
paths = {
'claude_desktop': os.path.expanduser('~/.config/claude/claude_desktop_config.json'),
'cursor': os.path.expanduser('~/.cursor/mcp.json'),
'project': '.mcp.json',
}
for name, path in paths.items():
if os.path.exists(path):
with open(path) as f:
data = json.load(f)
servers = data.get('mcpServers', {})
configs[name] = {
'path': path,
'servers': servers,
'server_count': len(servers),
}
return configs
configs = load_mcp_configs()
for name, cfg in configs.items():
print(f'{name} ({cfg["path"]}):')
print(f' Servers: {cfg["server_count"]}')
for server_name in cfg['servers']:
print(f' - {server_name}')Étape 2: Estimer les coûts en jetons par serveur
Comptez les jetons dans les descriptions d'outils avec tiktoken. Chaque outil a un nom, une description et un schéma de paramètres qui consomment tous des jetons.
Python
import tiktoken
def count_tokens(text: str, model: str = 'gpt-4o') -> int:
enc = tiktoken.encoding_for_model(model)
return len(enc.encode(text))
# Typical tool description sizes (measured from real MCP servers)
SERVER_ESTIMATES = {
'filesystem': {'tools': 11, 'avg_desc_tokens': 180},
'github': {'tools': 25, 'avg_desc_tokens': 250},
'slack': {'tools': 12, 'avg_desc_tokens': 200},
'postgres': {'tools': 8, 'avg_desc_tokens': 220},
'memory': {'tools': 5, 'avg_desc_tokens': 150},
'brave-search': {'tools': 2, 'avg_desc_tokens': 300},
'scavio': {'tools': 3, 'avg_desc_tokens': 250},
'tavily': {'tools': 2, 'avg_desc_tokens': 280},
'puppeteer': {'tools': 8, 'avg_desc_tokens': 200},
'sequential-thinking': {'tools': 1, 'avg_desc_tokens': 400},
}
def estimate_server_tokens(server_name: str) -> int:
"""Estimate token cost for a known MCP server."""
if server_name in SERVER_ESTIMATES:
est = SERVER_ESTIMATES[server_name]
return est['tools'] * est['avg_desc_tokens']
return 5 * 200 # Default: 5 tools x 200 tokens
# Audit all servers
print('MCP Server Token Audit')
print('=' * 50)
total = 0
for name, est in sorted(SERVER_ESTIMATES.items(), key=lambda x: x[1]['tools'] * x[1]['avg_desc_tokens'], reverse=True):
tokens = est['tools'] * est['avg_desc_tokens']
total += tokens
print(f'{name:25s} {est["tools"]:3d} tools x {est["avg_desc_tokens"]:3d} tok = {tokens:5,} tokens')
print(f'{"TOTAL":25s} {"":>18s} {total:5,} tokens')Étape 3: Calculer le coût des jetons de description d'outils
Convertissez les comptes de jetons en coûts réels en dollars. Avec la tarification actuelle des LLM, les jetons de description d'outils s'accumulent rapidement lorsque vous effectuez des centaines de requêtes par jour.
Python
def cost_analysis(servers: list[str], requests_per_day: int = 100) -> dict:
"""Calculate the real cost of MCP tool description tokens."""
total_tokens = sum(estimate_server_tokens(s) for s in servers)
# LLM input token pricing (2026 rates)
pricing = {
'gpt-4o': 2.50 / 1_000_000,
'gpt-4o-mini': 0.15 / 1_000_000,
'claude-sonnet-4': 3.00 / 1_000_000,
'claude-opus-4': 15.00 / 1_000_000,
}
results = {'total_tokens_per_request': total_tokens, 'servers': servers}
print(f'Servers: {", ".join(servers)}')
print(f'Token overhead per request: {total_tokens:,}')
print(f'\nMonthly cost at {requests_per_day} requests/day:')
for model, rate in pricing.items():
daily = total_tokens * requests_per_day * rate
monthly = daily * 30
print(f' {model:25s} ${monthly:6.2f}/month')
results[model] = monthly
return results
# Full global config: all servers
print('GLOBAL CONFIG (all servers):')
full = cost_analysis(list(SERVER_ESTIMATES.keys()), requests_per_day=200)
print()
# Scoped config: just Scavio
print('SCOPED CONFIG (Scavio only):')
scoped = cost_analysis(['scavio'], requests_per_day=200)
print()
print(f'Monthly savings with Claude Sonnet: '
f'${full.get("claude-sonnet-4",0) - scoped.get("claude-sonnet-4",0):.2f}')Étape 4: Générer des recommandations d'optimisation
Analysez votre configuration actuelle et recommandez quels serveurs limiter par projet. Identifiez les serveurs qui pourraient être remplacés par Scavio pour une consolidation.
Python
def optimize_mcp(current_servers: list[str]) -> dict:
"""Generate MCP optimization recommendations."""
# Servers that Scavio can replace
replaceable = {
'brave-search': 'Scavio covers web search + 5 more platforms',
'tavily': 'Scavio provides similar web search at $0.005/credit',
'google-search': 'Scavio includes Google search',
}
recommendations = []
tokens_before = sum(estimate_server_tokens(s) for s in current_servers)
new_servers = []
scavio_added = False
for s in current_servers:
if s in replaceable:
recommendations.append({
'action': 'REPLACE',
'server': s,
'reason': replaceable[s],
'token_savings': estimate_server_tokens(s),
})
if not scavio_added:
new_servers.append('scavio')
scavio_added = True
else:
new_servers.append(s)
tokens_after = sum(estimate_server_tokens(s) for s in new_servers)
print('MCP Optimization Recommendations')
print('=' * 50)
for rec in recommendations:
print(f'REPLACE {rec["server"]} -> Scavio')
print(f' Reason: {rec["reason"]}')
print(f' Token savings: {rec["token_savings"]:,}')
print(f'\nBefore: {tokens_before:,} tokens ({len(current_servers)} servers)')
print(f'After: {tokens_after:,} tokens ({len(new_servers)} servers)')
print(f'Saved: {tokens_before - tokens_after:,} tokens/request')
return {'before': tokens_before, 'after': tokens_after, 'recommendations': recommendations}
optimize_mcp(['filesystem', 'github', 'brave-search', 'tavily', 'slack', 'memory'])Exemple Python
Python
import json, os
SERVER_TOKEN_EST = {
'filesystem': 1980, 'github': 6250, 'slack': 2400,
'brave-search': 600, 'scavio': 750, 'tavily': 560,
'memory': 750, 'postgres': 1760,
}
def audit_mcp(config_path='.mcp.json'):
if not os.path.exists(config_path):
print(f'{config_path} not found')
return
with open(config_path) as f:
servers = json.load(f).get('mcpServers', {})
total = 0
for name in servers:
tokens = SERVER_TOKEN_EST.get(name, 1000)
total += tokens
print(f' {name}: ~{tokens:,} tokens')
print(f'Total: ~{total:,} tokens per request')
print(f'Monthly cost (Sonnet, 100 req/day): ${total * 100 * 30 * 3 / 1e6:.2f}')
audit_mcp()Exemple JavaScript
JavaScript
const fs = require('fs');
const TOKEN_EST = {
filesystem: 1980, github: 6250, slack: 2400,
'brave-search': 600, scavio: 750, tavily: 560,
memory: 750, postgres: 1760,
};
function auditMcp(configPath = '.mcp.json') {
if (!fs.existsSync(configPath)) return console.log('Config not found');
const servers = JSON.parse(fs.readFileSync(configPath, 'utf8')).mcpServers || {};
let total = 0;
for (const name of Object.keys(servers)) {
const tokens = TOKEN_EST[name] || 1000;
total += tokens;
console.log(` ${name}: ~${tokens.toLocaleString()} tokens`);
}
console.log(`Total: ~${total.toLocaleString()} tokens per request`);
console.log(`Monthly cost (Sonnet, 100 req/day): $${(total * 100 * 30 * 3 / 1e6).toFixed(2)}`);
}
auditMcp();Sortie attendue
JSON
MCP Server Token Audit
==================================================
github 25 tools x 250 tok = 6,250 tokens
filesystem 11 tools x 180 tok = 1,980 tokens
slack 12 tools x 200 tok = 2,400 tokens
postgres 8 tools x 220 tok = 1,760 tokens
scavio 3 tools x 250 tok = 750 tokens
TOTAL 19,310 tokens
MCP Optimization Recommendations
==================================================
REPLACE brave-search -> Scavio
Reason: Scavio covers web search + 5 more platforms
REPLACE tavily -> Scavio
Reason: Scavio provides similar web search at $0.005/credit
Before: 15,540 tokens (6 servers)
After: 13,130 tokens (5 servers)
Saved: 2,410 tokens/request