deepseek-harness-orb

Turns dsh from “open a browser” into “call it up from anywhere” — as a desktop orb.

What it solves

DeepSeek Harness is powerful but it is a Web UI: you switch windows and open a browser just to send one instruction. DeepSeek Orb makes dsh a floating ball parked at the edge of your screen — hover to expand, click to pin, drag past the edge to dock it into a thin gray tab, and talk to it whenever you want.

More importantly, it resolves the tension between “quick actions” and “long tasks” with a dual-track agent architecture:

• Foreground track · Computer Use: visible operations — opening apps, clicking buttons, filling forms, changing settings — run right in the frontmost window through 13 GUI tools (click, input_text, scroll, hotkey, drag, screenshot, open_app, and more), each step grounded in a fresh screenshot so the model always sees the current screen. Quick lookups like weather or news headlines are answered on the ball with web_search / web_fetch.
• Background track · Code Agent: complex work — digging through files, producing Word/PPT/Excel documents, building a website — is dispatched via code_agent to a background standard session. Dispatch returns immediately and you keep chatting. When the background session ends and the ball is idle, its completion summary returns to the ball automatically and the Computer Use Agent decides the next move: keep clicking, dispatch more background work, or wrap up.

There is no separate runtime task classifier — the Computer Use Agent inside the ball decides how every message is handled.

Other highlights

• The main window keeps the full dsh Web UI: session management, plugin market, model settings — you can still write code, edit files, and run commands. Background sessions show up in the dsh_orb folder in the sidebar where you can open, continue, or stop them; the ball’s Stop button only cancels the ball’s Computer Use session and never touches a background session.
• Each track picks its own model and thinking level (default DeepSeek-V41-Flash at Max thinking). The right-click menu also toggles coordinate encoding (millifraction on Windows, pixels on macOS) and the selection toolbar, and the avatar can be replaced with a custom GIF/PNG/WebP.
• Selection toolbar: select text in any app and get Search (Bing), Translate (result written into the current ball conversation), and Send to agent.
• Access chip: view-only / workspace edits / full access (full by default), applying to the ball’s commands and the background sessions it dispatches.
• No telemetry collected.

Who it’s for

• macOS / Windows users who want to direct an AI without leaving the window they are working in;
• developers already on DeepSeek Harness / the dsh CLI who want an always-on desktop entry point plus background long-task dispatch;
• anyone who needs AI to drive cross-app GUI work — filling forms, changing settings, repetitive clicking.

On Linux the main window works but there is no floating ball; on Windows, administrator-elevated windows refuse to be clicked or typed into; on macOS you must grant Screen Recording and Accessibility the first time you expand the panel. Read SAFETY.md before running.

DeepSeek Orb is an unofficial community project and is not affiliated with DeepSeek AI.

Getting started

git clone https://github.com/mini-yifan/deepseek-harness-orb.git
cd deepseek-harness-orb
pnpm install
pnpm run dev:desktop

Prerequisites: Node.js ^22.19.0 || >=24.0.0, pnpm 11.7.0, and a DeepSeek API key (save it in the main window under Settings → Models, or export DEEPSEEK_API_KEY=… first; keys live in ~/.dsh/.credentials.yaml, shared with the dsh CLI). Packaging commands and platform details are in MAC-RELEASE-PACK.md / UNSIGNED-MAC-PACK.md / UNSIGNED-WIN-PACK.md.