Citadel Predict forecasts AI agent token usage and cost before the agent runs — not after, like existing tools (Langfuse, Helicone, LangSmith). It analyzes your task and tool list, then uses a calibrated ML model to give you an expected token range, confidence level, and out-of-distribution risk flag — now converted into an estimated USD cost too.
Ships as a pip package + CLI, an MCP server for Claude Desktop/Code, and a hosted API. Tested on Linux, Windows, and macOS.
