Minimise what you send to and generate from an AI system, every token and every tool or MCP call, carries a real compute cost.
Easy to implement
High impact score
Input tokens, output tokens and tool-call payloads all have real energy behind them. This is worth being deliberate about with MCP (Model Context Protocol) tool calls in particular: each one typically returns a full schema definition plus a full response payload back into context, and chaining several calls across an agentic task compounds that quickly. Favour targeted retrieval over stuffing entire documents into context, trim conversation history once it’s no longer relevant to the task at hand, and treat every additional tool call as a cost to be justified rather than a free convenience.
EcoLogits tracks the energy consumption and emissions of GenAI API calls (including Anthropic and OpenAI) directly in your code.