CosTrack MCP — Agent Cost Control Layer

Lightweight MCP server that gives any AI agent or developer instant cost tracking, spend analysis, budget enforcement, and model cost optimization for LLM operations.

What It Does

CosTrack sits between simple price-lookup tools and full observability platforms. It's an MCP-native cost control layer — log costs, get reports, check budgets, compare models — all via tool calls.

Tools

Supported Models

• Anthropic: Claude Opus 4, Opus 4.6, Sonnet 4, Sonnet 4.6, Haiku 4
• OpenAI: GPT-4o, GPT-4o Mini, GPT-4 Turbo, o1, o1-mini
• Google: Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.0 Flash
• Meta: Llama 4 Scout, Llama 4 Maverick
• DeepSeek: DeepSeek V3, DeepSeek R1

Model names are automatically normalized — use aliases like sonnet, gpt-4o, haiku etc.

Quick Start

Connect via Claude Desktop / Claude Code

Add to your MCP server configuration:

{
  "mcpServers": {
    "costrack": {
      "url": "https://costrack-mcp.<your-subdomain>.workers.dev/mcp"
    }
  }
}

Example Usage

Log a cost event:

cost_log(model: "claude-sonnet-4", input_tokens: 1500, output_tokens: 800, agent_id: "my-agent")
→ { cost_usd: 0.0165, running_session_total: 1.23 }

Check budget:

budget_check(budget_usd: 100, scope: "all", period: "30d")
→ { current_spend_usd: 65.0, status: "PROJECTED_OVERAGE", projected_end_of_period_spend_usd: 130.1 }

Estimate before calling:

cost_estimate(model: "claude-opus-4", estimated_input_tokens: 5000, estimated_output_tokens: 2000)
→ { cost_per_call_usd: 0.225, alternatives: [{ model: "anthropic/claude-sonnet-4", savings_percent: 90.0 }] }

Features

• Model Normalization — sonnet, claude-sonnet-4, anthropic/claude-sonnet-4 all resolve to the same model
• Idempotency — Pass idempotency_key to prevent duplicate cost logging
• Hard Limit Signaling — Set hard_limit_usd to get alerts when spend exceeds threshold (fail-safe, not fail-stop)
• Price Snapshots — Each event stores the price at time of logging; historical costs never change
• Budget Projection — Predicts end-of-period spend based on daily average
• Alternative Suggestions — cost_estimate recommends cheaper models in same capability tier

Deployment

Prerequisites

• Node.js 18+
• Cloudflare account with Workers enabled
• Wrangler CLI (npm install -g wrangler)

Deploy

# Install dependencies
npm install

# Login to Cloudflare
wrangler login

# Create KV namespace
wrangler kv:namespace create COSTRACK_EVENTS
# Update wrangler.toml with the returned namespace ID

# Deploy
wrangler deploy

Local Development

npm run dev

Type Check

npm run build

Architecture

• Runtime: Cloudflare Workers (TypeScript)
• Storage: Hybrid — Durable Objects (real-time aggregations) + KV (raw event audit trail)
• Protocol: MCP (Model Context Protocol) over Streamable HTTP
• Pricing: Built-in table, single-file source of truth (src/pricing/pricing-table.ts)

Configuration

Edit wrangler.toml to set:

• COSTRACK_EVENTS KV namespace ID (Rich fills in after wrangler kv:namespace create)
• Production environment variables

Project Structure

costrack-mcp/
├── src/
│   ├── tools/           # 6 tool implementations
│   ├── pricing/         # Pricing data + model aliases
│   ├── normalize/       # Model name normalization
│   ├── storage/         # Durable Object + KV + DO client
│   ├── utils/           # Cost calculator, tier check
│   ├── types/           # Shared TypeScript types
│   ├── index.ts         # MCP server core
│   └── worker.ts        # Cloudflare Workers entry point
├── wrangler.toml        # CF Workers config
├── package.json
├── tsconfig.json
└── README.md

License

MIT