AI Meter

Local usage meter that turns AI coding-agent tokens into estimated electricity and water consumption.

Free· Free for individuals and companies; open source under a public GitHub repo.Evaluation

Best for

macOS developers and small engineering teams who use local AI coding agents daily and want a private, transparent sense of the energy and water cost of their sessions.

Skip if

Windows or Linux users, enterprises needing audited sustainability numbers, or teams that rely mostly on web-based ChatGPT/Claude rather than local CLI agents.

AI Meter is a macOS menu-bar app that watches the local log files written by AI coding agents (Claude Code, Codex, Cursor, OpenCode, Gemini CLI) and converts the tokens they consume into estimated electricity draw (kWh) and cooling water (liters). Everything runs on-device — the only outbound network call is a public update check — so no API keys, accounts, or telemetry leave the machine. Under the hood it reads each provider's local usage files, deduplicates overlapping events, aggregates by day and provider, and applies a research-backed conversion (default 0.39 kWh per million tokens, adjustable in a 0.20-0.75 range, plus a configurable PUE of 1.20 and site WUE of 0.45 L per IT-kWh) to translate raw counts into environmental impact. The UI presents interactive charts, per-provider breakdowns, and a live widget so developers can see, at a glance, how heavy today's agentic coding session was. It is aimed at engineers, tech leads, and sustainability-minded teams who want an honest, self-hosted view of the invisible externalities of AI-assisted development — not a marketing dashboard from the model vendor. Typical workflows: leaving the widget open during a Claude Code / Cursor session to watch the meter tick, comparing weekly consumption across agents to pick the most efficient tool for a task, or exporting rough per-project impact figures for an internal ESG or FinOps report. Because the calculations exclude training energy, embodied hardware, off-site grid mix, networking, and user-device power, the numbers are best treated as directional lower-bound estimates rather than audited measurements.

Editor's take

A refreshingly honest little utility: it tells you exactly which numbers it made up, which it measured, and which it deliberately ignored. That candor plus the local-only design makes it more useful than most 'AI sustainability' dashboards I've seen, even if the absolute figures are directional. The macOS-only, single-machine ceiling is the main thing keeping it from being a team tool.

— The AI Tool Bible editorial team

Pros

✅ 100% local processing — no accounts, no cloud, no API keys; reads existing provider files directly
✅ Covers the five most-used local coding agents in one place (Claude Code, Codex, Cursor, OpenCode, Gemini CLI)
✅ Transparent, user-adjustable math for kWh/token, PUE, and WUE rather than a black-box estimate
✅ Open source on GitHub, so the deduplication and parsing logic can be audited
✅ Free for individuals and companies with no gated tier
✅ Native macOS widget makes ambient monitoring effortless during long coding sessions
✅ Explicitly documents what it does NOT count (training, embodied hardware, grid carbon), which is rare for impact tools

Cons

⚠️ macOS only — no Windows or Linux build, which excludes a large slice of the developer audience
⚠️ Estimates are based on published research, not real telemetry from OpenAI/Anthropic/Google data centers
⚠️ Ignores training energy, embodied hardware, networking, and grid-mix carbon, so figures understate true impact
⚠️ Only tracks agents that write local usage files; hosted web chat sessions (chatgpt.com, claude.ai) are invisible
⚠️ No API or export hooks documented, which limits integration with FinOps or ESG dashboards
⚠️ Single-machine scope — no team roll-up view for a whole engineering org

Use cases

Tracking daily token usage across Claude Code and CursorEstimating electricity draw of an AI-assisted coding sessionEstimating cooling water footprint of AI coding workflowsComparing energy efficiency between competing coding agentsAmbient menu-bar monitoring of local LLM consumptionPersonal ESG or FinOps rough-order-of-magnitude reportingAuditing which agent burns the most tokens per feature shipped