MCP Server

@neurameter/mcp-server provides an MCP (Model Context Protocol) server that lets AI agents self-monitor their own cost and context usage in real-time.

Installation


npm install @neurameter/mcp-server

Usage


import { NeuraMeterMCPServer } from '@neurameter/mcp-server';
 
const server = new NeuraMeterMCPServer({
  apiKey: 'nm_xxx',
  projectId: 'proj_xxx',
});

Available Tools

The MCP server exposes these tools to connected agents:

`get_cost_summary`

Get cost summary for a time period.

Input:


{
  "agentName": "MyAgent",
  "period": "1h"
}

Output:


{
  "totalCost": 4.25,
  "totalEvents": 42,
  "topModels": [
    { "model": "gpt-4o", "cost": 3.80 },
    { "model": "gpt-4o-mini", "cost": 0.45 }
  ]
}

`check_context`

Analyze context window utilization for a set of messages.

Input:


{
  "messages": [...],
  "model": "gpt-4o"
}

Output:


{
  "utilizationPercent": 0.82,
  "estimatedInputTokens": 105000,
  "modelContextLimit": 128000,
  "breakdown": {
    "systemPromptTokens": 2500,
    "conversationTokens": 88000,
    "toolResultTokens": 14500
  },
  "suggestion": "Summarize conversation history to save ~60% of input tokens"
}

`check_budget`

Check remaining budget for the current period.

Input:


{
  "agentName": "MyAgent"
}

Output:


{
  "budgetLimit": 100.0,
  "spent": 45.20,
  "remaining": 54.80,
  "utilizationPercent": 0.452
}

`get_recommendations`

Get optimization recommendations based on recent usage patterns.

Input:


{
  "agentName": "MyAgent"
}

Output:


{
  "recommendations": [
    "Switch from gpt-4o to gpt-4o-mini for classification tasks — same accuracy, 94% cheaper",
    "Conversation history accounts for 78% of context — enable auto-summarization",
    "Tool result from search_web averages 12K tokens — consider truncating to 2K"
  ]
}

`log_optimization`

Log when an agent takes an optimization action (for tracking).

Input:


{
  "agentName": "MyAgent",
  "action": "summarized_history",
  "tokensBefore": 95000,
  "tokensAfter": 22000,
  "description": "Summarized 45 messages to 3 key points"
}

Why MCP?

The MCP server enables Layer 3: Autonomous Optimization — agents that:

Check their own cost mid-conversation
Analyze their own context utilization
Decide to summarize, switch models, or throttle
Report their optimizations back

This creates a feedback loop where agents self-optimize without human intervention.

Configuration with Claude Desktop

Add to your Claude Desktop MCP config:


{
  "mcpServers": {
    "neurameter": {
      "command": "npx",
      "args": ["-y", "@neurameter/mcp-server"],
      "env": {
        "NEURAMETER_API_KEY": "nm_xxx",
        "NEURAMETER_PROJECT_ID": "proj_xxx"
      }
    }
  }
}