I cut my LLM API costs by 71% — here's the open-source SDK I built

After my LangChain agent cost me $12 in one afternoon, I built AgentFuse.

What it does

Semantic caching — similar prompts return cached results without hitting the API (87.5% hit rate in benchmarks)

Per-run budget enforcement — hard cap spend per agent run before it blows up your bill

Zero infrastructure — no proxy server, just pip install and 2 lines of code

Install

pip install agentfuse-runtime

Works with

LangChain, CrewAI, LangGraph, OpenAI Agents SDK, MCP, Pydantic AI

Benchmarks

• 87.5% cache hit rate
• 71% cost reduction on repeated/similar prompts

GitHub: https://github.com/vinaybudideti/agentfuse
PyPI: https://pypi.org/project/agentfuse-runtime/

Would love feedback from anyone building agents.