Metrx MCP Server
Your AI agents are wasting money. Metrx finds out how much, and fixes it.
The official MCP server for Metrx — the AI Agent Cost Intelligence Platform. Give any MCP-compatible agent (Claude, GPT, Gemini, Cursor, Windsurf) the ability to track its own costs, detect waste, optimize model selection, and prove ROI.
Why Metrx?
Problem | What Metrx Does |
No visibility into agent spend | Real-time cost dashboards per agent, model, and provider |
Overpaying for LLM calls | Provider arbitrage finds cheaper models for the same task |
Runaway costs | Budget enforcement with auto-pause when limits are hit |
Wasted tokens | Cost leak scanner detects retry storms, context bloat, model mismatch |
Can't prove AI ROI | Revenue attribution links agent actions to business outcomes |
Quick Start
One-command install (Claude Desktop, Cursor, Windsurf)
{
"mcpServers": {
"metrx": {
"command": "npx",
"args": ["@metrxbot/mcp-server"],
"env": {
"METRX_API_KEY": "sk_live_your_key_here"
}
}
}
}Get your free API key at metrxbot.com/settings/security.
Remote HTTP endpoint
For remote agents (no local install needed):
POST https://metrxbot.com/api/mcp
Authorization: Bearer sk_live_your_key_here
Content-Type: application/jsonFrom npm
npm install @metrxbot/mcp-server23 Tools Across 10 Domains
Dashboard (2 tools)
Tool | Description |
| Total spend, call counts, error rates, agent breakdown, and optimization opportunities |
| All agents with status, category, cost metrics, and health indicators |
Optimization (4 tools)
Tool | Description |
| Compare costs across providers — find cheaper alternatives |
| Revenue per agent with confidence scores and ROI metrics |
| Token limit recommendations and overflow detection |
| Model switching recommendations based on cost, latency, quality |
Budgets (4 tools)
Tool | Description |
| Create monthly/daily budgets with hard, soft, or monitor enforcement |
| Update limits, frequency, or enforcement mode |
| All budgets with current spend vs. limits |
| Remove a budget (historical data preserved) |
Alerts (3 tools)
Tool | Description |
| Alert on cost overages, error rates, latency spikes, anomalies |
| Update thresholds, channels, enable/disable |
| Active alerts and current status per agent |
Experiments (2 tools)
Tool | Description |
| A/B test comparing two LLM models with traffic splitting |
| Statistical significance, cost delta, and recommended action |
Cost Leak Detector (2 tools)
Tool | Description |
| Find cost anomalies and waste across your fleet |
| Deep-dive into a specific anomaly with timeline and root cause |
Attribution (2 tools)
Tool | Description |
| Link agent actions to business outcomes for ROI tracking |
| Multi-source attribution report with confidence scores |
ROI & Reporting (2 tools)
Tool | Description |
| ROI report for tier upgrades based on usage patterns |
| Board-ready ROI audit report |
Alert Configuration (2 tools)
Tool | Description |
| Set cost/operational thresholds with email, webhook, or auto-pause |
| Predictive analysis — identify agents likely to fail before it happens |
Prompts
Pre-built prompt templates for common workflows:
Prompt | Description |
| Comprehensive cost overview — spend breakdown, top agents, optimization opportunities |
| Discover optimization opportunities — model downgrades, caching, routing |
| Scan for waste patterns — retry storms, oversized contexts, model mismatch |
Examples
"How much am I spending?"
User: What was my AI cost this week?
→ metrx_get_cost_summary(period_days=7)
Total Spend: $234.56 | Calls: 2,450 | Error Rate: 0.2%
├── customer-support: $156.23 (1,800 calls)
└── code-generator: $78.33 (650 calls)
💡 Switch customer-support from GPT-4 to Claude Sonnet: Save $42/week"Find me savings"
User: Am I overpaying for my agents?
→ metrx_get_provider_arbitrage(agent_id="agent_123")
Current: GPT-4 @ $15.20/1K calls
Alternative: Gemini 1.5 @ $6.80/1K calls (-55%)
Estimated Savings: $420/month"Test a cheaper model"
User: Test Claude 3.5 Sonnet against my GPT-4 setup
→ metrx_start_experiment(name="Claude Trial", agent_id="agent_123",
model_a="gpt-4", model_b="claude-3-5-sonnet", traffic_split=10)
Experiment started: 90% GPT-4, 10% Claude 3.5 Sonnet
Check back in 14 days for statistical significance.Configuration
Variable | Required | Description |
| Yes | Your Metrx API key (get one free) |
| No | Override API base URL (default: |
Rate Limiting
60 requests per minute per tool. For higher limits, contact support@metrxbot.com.
Development
git clone https://github.com/metrxbots/mcp-server.git
cd mcp-server
npm install
npm run typecheck
npm testContributing
See CONTRIBUTING.md for guidelines.
Links
Website: metrxbot.com
Docs: docs.metrxbot.com
npm: @metrxbot/mcp-server
Smithery: metrxbot/mcp-server
Support: support@metrxbot.com
License
MIT — see LICENSE.