llm-cost-watch

詢問 Claude 為什麼你的 LLM 帳單突然暴增。這是一個開源 FastMCP 伺服器,可分析功能層級的追蹤記錄、診斷多代理程式的 token 迴圈,並將簡單任務路由至 Haiku,最多可降低 94% 的 API 支出。

文件

llm-cost-watch

The Model Context Protocol (MCP) server for LLM spend & latency intelligence.
"Ask Claude why our LLM bill jumped."

Product Hunt MCP 2.x Python 3.11+ TrustMRR Verified Gumroad Flash Sale License: MIT


📺 90-Second Explainer Video Walkthrough

Watch the 90-Second Walkthrough

▶️ Click here to download / watch the full 1080p MP4 walkthrough (with voiceover)
Explaining why multi-agent swarms face compounding token spikes, how local-first MCP telemetry works, and how to cut LLM spend by 94% with automated micro-model routing.


What is llm-cost-watch?

Instead of clicking through complex web dashboards, llm-cost-watch lets you ask Claude directly inside Claude Desktop, Claude Code, or Antigravity:

You: "Claude, why did our LLM spend jump 40% this week, and where is our latency bottleneck?"

Claude: [Calls llm-cost-watch: cost_by_feature & slow_calls]
"Your 'knowledge_base_rag' pipeline accounts for 66.1% of total spend ($0.082 across 3 test runs) 
and is driving your p95 latency up to 3,120ms due to uncompressed 7k token context windows.

Meanwhile, 'customer_intent_classifier' is currently routed through GPT-4o for 14-token classification. 
Routing it to GPT-4o-mini will cut that feature's spend by 94% and reduce p95 latency to 320ms."

3 Core MCP Tools

  1. cost_by_feature(timeframe="7d"): Identifies which feature or workflow consumes the largest share of your budget.
  2. slow_calls(limit=5, min_latency_ms=1200): Pinpoints high-latency outliers (p50, p95, p99) and diagnoses the bottleneck (unbounded prompt payload, recursive tool loops, reasoning stalls).
  3. routing_suggestions(savings_threshold_pct=20): Analyzes short structured prompts running on expensive frontier models and calculates exact dollar savings from moving to Claude 3.5 Haiku or GPT-4o-mini.

Quick Setup (Under 5 Minutes)

1. Install & Clone

git clone https://github.com/rohith0707/llm-cost-watch.git
cd llm-cost-watch
pip install -r requirements.txt

2. Connect to Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "llm-cost-watch": {
      "command": "python",
      "args": ["-m", "tools.llm-cost-watch.server"]
    }
  }
}

3. Log Calls in Your Code (One Line)

from logger import watch_llm

@watch_llm(feature="customer_intent_classifier", task_type="classification", model="gpt-4o")
def classify_intent(message: str) -> str:
    # Your normal OpenAI / Anthropic call
    return client.chat.completions.create(...)

Test Run

python test_server.py

Commercial Licensing & Architecture Audits