@reaatech/llm-cost-telemetry-mcp

Status: Pre-1.0 — APIs may change in minor versions. Pin to a specific version in production.

MCP (Model Context Protocol) server that exposes LLM cost telemetry as MCP tools. Provides a three-layer tool architecture for recording cost spans, querying aggregated costs, and enforcing budget limits — all consumable by MCP clients like Claude Desktop, Cursor, and custom agent frameworks.

Installation

npm install @reaatech/llm-cost-telemetry-mcp @modelcontextprotocol/sdk
# or
pnpm add @reaatech/llm-cost-telemetry-mcp @modelcontextprotocol/sdk

Feature Overview

• Three-layer MCP tool architecture — atomic span operations, stateful aggregation, and budget enforcement
• Cost span recording and retrieval — record spans by ID, flush buffered spans
• Multi-dimensional queries — query by tenant, feature, or route with time windowing
• Budget checks — pre-call budget gating with cascading alert thresholds
• Cost calculation on flush — automatically calculates cost from token counts via the calculator engine
• Pluggable span handler — register a callback to forward spans to exporters or custom pipelines

Quick Start

import { createCostTelemetryServer } from "@reaatech/llm-cost-telemetry-mcp";
 
const server = createCostTelemetryServer({
  onSpanFlush: (spans) => {
    for (const span of spans) {
      console.log(`Cost: $${span.costUsd} (${span.provider}/${span.model})`);
    }
  },
});
 
// Connect via stdio — ready for MCP clients
await server.connect(transport);

API Reference

createCostTelemetryServer(options?): Server

Creates a fully configured MCP server with 10 tools across three layers.

import { createCostTelemetryServer, type MCPServerOptions } from "@reaatech/llm-cost-telemetry-mcp";
 
const server = createCostTelemetryServer({
  collectorOptions: { maxBufferSize: 1000, flushIntervalMs: 60000 },
  aggregatorOptions: { dimensions: ["tenant", "feature", "route"] },
  budgetConfig: {
    tenants: { "acme-corp": { daily: 100, monthly: 2000 } },
  },
  onSpanFlush: (spans) => { /* forward to exporters */ },
});

MCPServerOptions

Three-Layer Tool Architecture

Layer 1: cost.span.* — Atomic Operations

Fast, stateless operations for recording individual cost spans:

Example: Record a cost span via MCP tool call

{
  "name": "cost.span.record",
  "arguments": {
    "provider": "openai",
    "model": "gpt-4",
    "inputTokens": 150,
    "outputTokens": 45,
    "tenant": "acme-corp",
    "feature": "chat-support"
  }
}

On flush, the server automatically calls calculateCost() to fill in costUsd from the token counts.

Layer 2: cost.aggregate.* — Aggregation

Stateful operations for cost aggregation by dimensions:

Example: Query costs by tenant

{
  "name": "cost.aggregate.by_tenant",
  "arguments": {
    "tenant": "acme-corp",
    "period": "day"
  }
}

Layer 3: cost.budget.* — Budget Management

Opinionated operations for budget enforcement:

Example: Check budget before API call

{
  "name": "cost.budget.check",
  "arguments": {
    "tenant": "acme-corp",
    "estimatedCost": 5.00
  }
}

Returns { withinBudget: true, dailyPercentage: 42, monthlyPercentage: 3, triggeredAlerts: ["log"] }.

Usage Patterns

Integration with MCP Clients

import { createCostTelemetryServer } from "@reaatech/llm-cost-telemetry-mcp";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 
const server = createCostTelemetryServer({
  onSpanFlush: (spans) => {
    // Forward spans to AWS CloudWatch, GCP Cloud Monitoring, etc.
  },
});
 
const transport = new StdioServerTransport();
await server.connect(transport);

Custom Aggregation Dimensions

const server = createCostTelemetryServer({
  aggregatorOptions: {
    dimensions: ["tenant", "feature", "route", "provider", "model"],
    timeWindows: ["hour", "day", "week", "month"],
  },
});

Budget Enforcement Flow

MCP Client                    MCP Server
    |                             |
    | cost.budget.check           |
    | { tenant, estimatedCost }   |
    |---------------------------->|
    |                             | → BudgetManager.check()
    | { withinBudget, ... }       |
    |<----------------------------|
    |                             |
    | cost.span.record            |
    | { provider, model, ... }    |
    |---------------------------->|
    | { span_id }                 | → CostCollector.add()
    |<----------------------------|

Related Packages

• @reaatech/llm-cost-telemetry — Core types and utilities
• @reaatech/llm-cost-telemetry-calculator — Cost calculation engine
• @reaatech/llm-cost-telemetry-aggregation — Aggregation and budgets
• @reaatech/llm-cost-telemetry-exporters — Telemetry exporters

License

MIT