A new open-source tool called llm-cost-watch is attempting to solve one of the most persistent headaches in the AI development lifecycle: opaque LLM billing. Released as a Model Context Protocol (MCP) server, the tool allows developers to query their cloud spend and latency metrics using natural language directly within Claude Desktop, Claude Code, or Antigravity. Instead of navigating complex web dashboards, users can simply ask, "Why did our LLM spend jump 40% this week?" and receive specific, data-driven answers.

Three Core Diagnostic Tools

The server exposes three distinct MCP tools designed to break down cost and performance bottlenecks. The cost_by_feature tool identifies which workflows are consuming the largest share of the budget over a specified timeframe. Meanwhile, slow_calls pinpoints high-latency outliers by analyzing p50, p95, and p99 metrics, diagnosing issues like unbounded prompt payloads or recursive tool loops. Finally, routing_suggestions analyzes short, structured prompts currently running on expensive frontier models, calculating exact dollar savings from migrating them to cheaper alternatives like Claude 3.5 Haiku or GPT-4o-mini.

From Dashboard to Conversation

The tool's value proposition lies in its ability to contextualize raw data into actionable engineering advice. In a demonstration, the tool identified that a knowledge_base_rag pipeline accounted for 66.1% of total spend due to uncompressed 7k token context windows driving p95 latency to 3,120ms. It simultaneously flagged a customer_intent_classifier running on GPT-4o for trivial 14-token tasks, suggesting a move to GPT-4o-mini to cut that feature's spend by 94%. This level of granular, conversational insight eliminates the manual correlation work typically required to optimize LLM applications.

Setup and Licensing

Installation is designed for rapid deployment, requiring users to clone the repository, install dependencies via pip, and add a configuration block to their claude_desktop_config.json. Instrumentation involves a single decorator line in the codebase: @watch_llm, which tags calls with feature names, task types, and model identifiers. While the core functionality is open-source under an MIT license, the developer, Rohith Balsa, offers a $2,500 Enterprise SLA and 14-Day Concurrency/Cost Audit for organizations needing deeper architectural support.

Key Takeaways

  • llm-cost-watch integrates with Claude Desktop, Claude Code, and Antigravity via the Model Context Protocol (MCP) to provide natural language cost analysis.
  • The tool offers three core diagnostic functions: cost_by_feature for budget allocation, slow_calls for latency outliers, and routing_suggestions for model optimization.
  • Setup requires cloning the repo, installing dependencies, adding a config block, and using a single @watch_llm decorator in the code.
  • The project is MIT-licensed open-source, with optional paid Enterprise SLA and audits available from the creator.

The Bottom Line

This is a pragmatic, necessary evolution in LLM observability. As model costs become a line-item expense, tools that translate raw usage data into immediate engineering actions will define the next generation of AI infrastructure.