Green DSH Desktop

An open-source desktop app for DeepSeek Harness (DSH): get AI work done while seeing exactly how much energy, carbon, and money each conversation costs.

Whether you run local models or connect to a cloud API (DeepSeek / OpenRouter / a self-hosted relay…), it calculates and displays the energy consumption, carbon emissions, and cost of every request, and warns you when spending or emissions spike.

English · 中文

---

What It Does

1. DSH Desktop, Ready Out of the Box

• Ships with a Node.js 24 runtime and a complete DSH distribution — ready to use after installation, no Node / pnpm / Python required;
• The full DSH chat experience: model management, sessions, tool calls, skills, sub-agents, and workflows are all included;
• Windows installer + portable edition (unzip and run).

2. Energy / Carbon Console (Main Window)

Page	What You Can See
Console	Local carbon, local energy (estimated and measured side by side), cloud carbon (public-coefficient estimate), API costs saved; last-30-days carbon and per-task-round carbon bar charts; real-time hardware gauges
Sessions	Per-session tokens, energy, carbon, electricity cost, and API costs saved
Hardware Monitor	Full NVIDIA GPU telemetry (memory / utilization / temperature / power, multi-GPU support); iGPU as extended monitoring; CPU / memory curves
Alerts	Overspending and anomaly warnings, see below
Data Sources	Connect OpenRouter or your own relay to pull real spend and token usage
Reference Data	Source and year of every coefficient, all verifiable
Carbon Ledger	Per-step JSONL details, measured-energy column, one-click CSV export
Logs / Settings	Harness output; baseline settings for carbon intensity / electricity price / API pricing

3. Two Metering Paths (Core Design)

• Local model → measured: once an NVIDIA GPU is detected, it continuously samples power draw and integrates over time to get real energy consumption, then attributes it to each request step by time; when a per-model coefficient is needed, one-click calibration yields the machine's J/token;
• Cloud API → public-coefficient estimate: the cloud only exposes tokens, so token→energy is always computed with public authoritative coefficients (never a local calibration extrapolated to the cloud).

4. Alerts: Warn You Before Money and Carbon Run Away

• Threshold-based: per-session energy / cost, last-1-hour / 24-hour cost, relay's daily average spend;
• Anomaly-based: single-step energy outlier (robust z-score), context re-reading (input:output ratio too high — wasteful repeated re-reads of context), idle / retry loops (multiple consecutive steps with almost no output), low cache hit rate;
• Severe alerts also fire a desktop notification.

5. In-Chat Dashboard

A carbon card is pinned to the chat window's top-right: this session's carbon / energy / electricity cost, API costs saved, local vs cloud, per-round energy bars, and cumulative totals.

6. Local Small-Model Support Out of the Box

Includes a Qwen2.5-0.5B (CPU inference) setup script and provider configuration — start the service with a single command and experience the full "local model + measured metering" loop; to use a larger model, just change the model path in one line.

---

Screenshots

Console Overview	In-Chat Dashboard

Hardware Monitor (NVIDIA Multi-GPU)	Alerts

Session Details	Data Sources (Relay)

Reference Data (Coefficients & Sources)	Carbon Ledger Details

---

Download & Install

Download from Releases:

File	Description
Green DSH Desktop Setup x.x.x.exe	Windows installer (optional install directory, desktop shortcut)
Green DSH Desktop-x.x.x-win.zip	Portable edition, unzip and run

The installer is not code-signed; on first launch, Windows SmartScreen will block it: More info → Run anyway. To upgrade, just run the new installer — sessions / credentials / ledger / settings are preserved automatically.

Quick Start (Three Steps)

1. Install and open → on first launch you'll be prompted to configure a model (enter an API Key in the DSH settings page, or use a local model);
2. Click "Open Chat" to work → chat and write code normally, nothing else to configure;
3. Return to the console → carbon, energy, cost, round charts, and hardware curves are all updating in real time; fine-tune electricity price / API pricing / carbon intensity in "Settings" and the numbers reflect your actual baseline.

To try a local model:

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

Then select "Local llama.cpp (Qwen2.5-0.5B)" in the app (you'll need to download the model first — see the comments inside the script).

---

Data & Methodology (Brief)

• Local: token usage → calibration coefficient → energy (measured value when NVIDIA is available) → grid carbon intensity → carbon;
• Cloud: token usage → public coefficient → energy → carbon;
• Cost: API list price vs local electricity cost.

All coefficients carry source and year:

Category	Value	Source
China grid	0.5777 kgCO₂e/kWh (2024)	MEE 《2024 Electric Power Carbon Footprint Factor》
US grid	0.3497 kgCO₂e/kWh	US EPA eGRID 2023
Cloud per-token	0.0006 Wh / output token	Epoch AI 2025
Cloud per-request	0.24 Wh / 0.03 gCO₂e	Google Cloud 2025-08-21
Local calibration	H20 request-level fit (R² ≥ 0.998)	Experiments in this project

Boundary notes: without GPU measurement, local values are token-profile estimates; cloud values are public-coefficient estimates; carbon = energy × grid intensity (engineering estimate, not certified).

---

Build from Source

git clone <本仓库>
cd green-dsh-desktop
npm install

# 1) 打包 green-meter / 仪表盘插件（需要 DSH 源码 checkout，可用 DSH_CHECKOUT 指定路径）
npm run vendor:plugins

# 2) 组装内置 harness（npm 安装 DSH 发行版 + 本地插件）
npm run pack:harness

# 3) 放入 Node 运行时（复制本机 node.exe 到 resources/runtime/）

# 4) 运行 / 打包
npm start
npm run dist

---

Limitations & Notes

• Currently developed and verified primarily on Windows; Electron is cross-platform, but the macOS / Linux build paths are unverified;
• Local measured metering depends on nvidia-smi (NVIDIA GPU); without NVIDIA, local values fall back to estimates, and iGPU monitoring is extended only (utilization only);
• Chat data (sessions / credentials / ledger) is stored in resources/harness/dsh-home under the install directory — please back up before uninstalling (upgrades preserve it automatically).

Credits

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

License

MIT