Launching today
llm-cost-watch
Ask Claude why your LLM bill jumped
1 follower
Ask Claude why your LLM bill jumped
1 follower
An open-source MCP server that lets Claude analyze feature-level trace logs, diagnose latency bottlenecks, and pinpoint automated model downgrades (Sonnet to Haiku) to cut LLM spend by up to 94% with zero quality loss.
Launch Team
LoomShow, don't just tell with Loom
Promoted
Hey Product Hunt! I'm Rohith Balsa, Core LLM & Agentic Systems Architect at KPMG India.
I built llm-cost-watch because multi-agent systems and parallel tool-calling loops face a hidden issue: compounding token spikes. When an agent loops or retries a reasoning chain, token burn compounds rapidly before generic APMs trigger alerts.
llm-cost-watch is a lightweight, local-first Model Context Protocol (MCP) server that hooks directly into Claude Desktop, Cursor, and custom agent swarms:
- Zero-Latency Telemetry: Inspect token spikes and latency bottlenecks in real-time before commits.
- Deterministic Cost Attribution: Pinpoint per-feature LLM spend across parallel workers.
- Automated Model Routing: Safely route high-frequency tasks to Haiku or GPT-4o-mini without quality loss.
- 100% open-source under Apache-2.0: https://github.com/rohith0707/ll...
Would love your thoughts, questions, and feedback on deterministic cost controls for production agent swarms!
For anyone interested in a quick visual teardown, we also put together a 90-second explainer video walkthrough showing how multi-agent pipelines burn tokens and how local-first FastMCP telemetry fixes it:
https://github.com/rohith0707/ll...