dsh-green-meter

An open-source desktop app for DeepSeek Harness (DSH): use AI as usual, and see exactly how much electricity, carbon, and money every conversation costs.

Whether you run local models or cloud APIs (DeepSeek / OpenRouter / your own gateway…), it accounts for the energy, carbon, and cost of every request, keeps them visible, and warns you when things get expensive.

中文 · English

What it does

1. A DSH desktop app that just works

• Ships with a bundled Node.js 24 runtime and the full DSH distribution — install and run, no Node / pnpm / Python needed;
• The complete DSH chat experience: model management, sessions, tool calls, skills, subagents, workflows;
• Windows installer + portable zip.

2. The energy / carbon console (main window)

3. Two accounting paths (the core design)

• Local models → measured: when an NVIDIA GPU is present, power is sampled continuously and integrated into real energy, attributed to each request step by time; a one-click calibration derives your machine's own J/token per model;
• Cloud APIs → published coefficients: the cloud only exposes tokens, so the token→energy relation always uses published, authoritative coefficients (never a local calibration projected onto the cloud).

4. Alerts: know when you're burning money and carbon

• Thresholds: per-session energy / cost, rolling 1-hour / 24-hour cost, relay daily average;
• Anomalies: per-step energy outliers (robust z-score), context thrash (huge input-to-output ratio — re-reading context with no useful output), retry loops (many steps with almost no output), low cache-hit rate;
• Critical alerts also raise a desktop notification.

5. In-chat dashboard

A carbon card pinned to the top-right of the chat window: session carbon / energy / electricity cost, API cost saved, local-vs-cloud comparison, per-turn energy bars, and all-session totals.

6. Local small-model support out of the box

Ships with a Qwen2.5-0.5B (CPU inference) launcher script and provider config — one command starts the server and you get the full "local model + measured accounting" loop; swap in a bigger model by changing one path.

Screenshots

Console overview	In-chat dashboard

Hardware (multi-GPU NVIDIA)	Alerts

Sessions	Data sources (relay)

Reference data (coefficients & sources)	Ledger detail

Download & install

Grab from Releases:

The installer is unsigned, so SmartScreen will interrupt the first launch: More info → Run anyway. Upgrades keep your sessions / credentials / ledger / settings automatically.

Quick start (three steps)

1. Install and open — first run helps you configure a model (paste an API key in the DSH settings page, or use a local model);
2. Click "Open Chat" and work — chat and code as usual, nothing else to change;
3. Back to the console — carbon, energy, cost, per-turn chart, and hardware curves refresh live; tune electricity price / API prices / grid intensity in Settings so the numbers match your reality.

To try a local model:

powershell -ExecutionPolicy Bypass -File scripts\start-local-model.ps1

then pick "本地 llama.cpp（Qwen2.5-0.5B）" in the app (download the model first — see the notes inside the script).

Data & method (in short)

• Local: token usage → calibration coefficients → energy (measurement wins when an NVIDIA GPU is present) → grid carbon intensity → carbon;
• Cloud: token usage → published coefficients → energy → carbon;
• Cost: API list price vs local electricity.

Every coefficient carries its source and year:

Honest boundaries: without GPU measurement, local figures are token-profile estimates rather than hardware measurements; cloud figures are estimates from published coefficients, not vendor measurements; carbon = energy × grid intensity is an engineering estimate, not certified accounting.

Build from source

git clone <this repo>
cd green-dsh-desktop
npm install

# 1) Pack the green-meter / dashboard plugins (needs a DSH source checkout; set DSH_CHECKOUT to point at it)
npm run vendor:plugins

# 2) Assemble the bundled harness (npm-installs the DSH distribution plus local plugins)
npm run pack:harness

# 3) Drop in the Node runtime (copy your node.exe into resources/runtime/)

# 4) Run / package
npm start
npm run dist

Limitations

• Developed and verified on Windows; Electron is cross-platform but macOS / Linux build paths are unverified;
• Local measurement relies on nvidia-smi (NVIDIA GPUs); without one, local figures fall back to estimates and integrated GPUs are monitored as an extension (utilization only);
• App data (sessions / credentials / ledger) lives in resources/harness/dsh-home inside the install directory — back it up before uninstalling (upgrades preserve it automatically).

Credits

• DeepSeek Harness and Cordis — this app is a plugin-based desktop distribution of DSH;
• green-meter (this project's energy-accounting plugin) provides the token-profile calibration;
• llama.cpp — local inference;
• Electron — the desktop shell;
• Energy / carbon reference data from published disclosures by MEE (China), US EPA eGRID, Epoch AI, and Google Cloud.

License

MIT