AgentFuse

AgentFuse is a Model Context Protocol server that tracks token spend across AI agent sessions in real-time. It warns at 80% of budget and fires a hard HALT — not a soft warning — at 100%, preventing runaway agent loops from burning money. Supports 18 built-in model pricing tables (GPT-4o, Claude Opus/Sonnet/Haiku, Gemini, DeepSeek, and more). Drop in 4 lines of MCP config and your agent cannot spend more than $X per session.