Seeki

A tiny DeepSeek-powered desktop companion that lives on your screen and helps you get things done.

English · 简体中文

Seeki is a pixel pet that floats on your desktop and keeps a DeepSeek Harness agent one double-click away. It's a Desktop Pet × DeepSeek × Agent — persistent conversations organized by workspace, per-task model switching, and a character that actually reacts while your agent is working.

Seeki is an independent community project. It is not an official DeepSeek product.

Why Seeki?

You can already talk to DeepSeek in a browser tab. Seeki adds the layer a chat UI can't:

Every one of these maps to something Seeki actually does today — no marketing filler.

Features

🐾 Lives on your desktop

Seeki is a transparent, frameless, always-on-top window — a 2D pixel character that stays on your screen instead of hiding in a tab.

💬 Persistent conversations

Double-click Seeki to chat. Conversations are persistent and multi-turn, organized into workspaces, with a conversation switcher to jump back into any previous thread.

🧠 A DeepSeek Harness agent behind the pet

Seeki is the desktop face of a DeepSeek Harness agent. The head bubble reflects the agent's live status:

🔀 Per-task model switching

Switch between your configured DeepSeek models straight from the chat window. The switch applies to the next message only and never interrupts a running turn; a new conversation inherits the current model.

🎮 A character, not a static icon

Idle breathing, happy, walking (drag), and sleeping animations — plus drag-direction switching (left / right / up / down with mirroring) and click / idle-timeout interactions.

⚙️ Config-driven, no code edits

Status bubbles, animations, frame rates, and triggers all live in pet.config.json, editable through a built-in /pet/settings page — including uploading, replacing, and deleting animation frames.

Screenshots

💬 Conversation

🗂️ Workspace & model switching

🎮 Character states

Quick Start

Prerequisites

• Node.js ≥ 22 and pnpm
• A clone of deepseek-harness with pnpm install run once — Seeki runs as a plugin inside it
• A DeepSeek API key configured for the harness (see Model Configuration)

# 1. Clone both repos
git clone https://github.com/deepseek-ai/deepseek-harness.git
git clone https://github.com/BenjaminSHI4008/deepseek-pet.git
cd deepseek-harness && pnpm install

# 2. Install Seeki (plugin + a desktop icon)
cd ../deepseek-pet
bash scripts/install.sh

# Windows (PowerShell)
powershell -ExecutionPolicy Bypass -File scripts\install.ps1

3. Run — double-click Seeki.app (macOS) / Seeki.lnk (Windows). The icon launches DeepSeek Harness in the background and wakes Seeki.

The installer auto-detects your deepseek-harness checkout (a sibling directory, ~/deepseek-harness, or ~/Projects/deepseek-harness) and asks if it can't find it.

Installation

Two ways to run, plus a standalone offline mode.

A. One-click desktop icon (recommended)

scripts/install.sh (macOS) / scripts/install.ps1 (Windows) installs the plugin and creates a pseudo-executable desktop icon — it contains only a launch command, no bundled runtime.

• Start: double-click → starts the harness in the background → wakes Seeki.
• Exit: right-click the pet → 「退出桌宠」. Quitting Seeki does not stop the harness (it keeps running in the background; the next double-click just re-wakes the pet).

B. Run with DeepSeek Harness manually

cd <deepseek-pet>
bash scripts/install-harness-plugin.sh   # install the plugin into the web profile

cd <deepseek-harness>
pnpm dsh web                            # Seeki auto-launches with the harness

C. Standalone pet (offline, no agent)

cd <deepseek-pet>
npm install
npm start

This runs only the pet. Without a connected harness the bubble shows Offline.

Electron downloads its binary on npm install. On slow networks in mainland China, run export ELECTRON_MIRROR="https://npmmirror.com/mirrors/electron/" first.

Model Configuration

Seeki doesn't hardcode models. The chat window's model list comes from your DeepSeek Harness's configured providers (llm.models), so it always reflects what your harness actually advertises (for example DeepSeek-V4-Pro and DeepSeek-V4-Flash on the deepseek-official route).

• API key — configure it for the harness, not for Seeki: set DEEPSEEK_API_KEY in the environment, in a .env file at the harness root, or via the harness's credentials store (~/.dsh/.credentials.yaml).
• Default model — the harness's own default. Seeki doesn't impose one.
• Switch models — use the model dropdown in the chat title bar. It affects the next message only; new conversations inherit the current selection; unavailable models show as disabled.

Full pet.config.json reference: docs/config.md.

How It Works

flowchart TB
    U[You] -->|double-click · type| P[Seeki Desktop UI<br/>Electron · transparent window · pixel sprite]
    P -->|WebSocket /api/pet.ws| S[dsh-pet-status plugin<br/>inside DeepSeek Harness]
    S -->|session.prompt · session.models| A[DeepSeek Harness Agent<br/>workspaces · sessions · tools]
    A -->|provider route| M[DeepSeek Model Provider]
    M --> A
    A --> S -->|status · assistant text| P

1. The Electron pet renders the character and the chat window.
2. A small Cordis plugin (dsh-pet-status) runs inside DeepSeek Harness, folding agent activity into statuses and streaming assistant text over a local WebSocket.
3. The harness agent handles sessions, workspaces, and tools; Seeki is its desktop face.

Architecture and extension guide: docs/development.md.

Project Structure

harness-plugin/     dsh-pet-status plugin — status broadcast, pet manager, chat bridge
Deepseek/           pixel sprite assets (animations, 8-direction stills, raw sources)
assets/             bundled font (and the generated launcher icon)
docs/               config reference + developer guide
scripts/            installers + sprite normalization
pet.config.json     all pet behavior — statuses, actions, character states
main.cjs            Electron main process (window, drag, WebSocket client)
renderer.js         pixel sprite state machine (config-driven)
chat.html           chat window UI (workspace, conversation, model switching)

Development

npm install
npm start        # run the pet standalone (offline) while iterating on the UI

Pet behavior is entirely config-driven — editing pet.config.json (or using /pet/settings) changes bubbles, animations, and triggers without touching render code. See docs/development.md for how to add an action, a mouse state, or a drag direction.

Roadmap

• Desktop pet (transparent always-on-top window, pixel state machine)
• Agent status reflection (received / running / completed / terminated / offline)
• Chat: workspaces, conversation history, model switching, cancel
• Visual settings page (statuses / actions / frame upload & delete)
• One-click desktop launcher (Seeki.app / Seeki.lnk)
• Click-through (transparent areas don't capture the mouse)
• Position memory, edge snapping, 8-direction walking
• Standalone packaging (bundled harness runtime)

Contributing

Issues, feature requests, and pull requests are welcome.

License

MIT © 2026 BenjaminSHI4008

Character sprite assets under Deepseek/ were generated with PixelLab; redistribution terms are per PixelLab's license.