KVDesk

Versatile Almost Local, Eventually Reasonable Assistant

Desktop AI Assistant with Local Model Support

https://github.com/user-attachments/assets/a8c54ce0-2fe0-40c3-8018-026cab9d7483

✨ Features

Core Capabilities

• ✅ Task Planning — visual todo panel with progress tracking, persisted per session
• ✅ Scheduled Tasks — create reminders and recurring tasks with auto-execution
• ✅ OpenAI SDK — full API control, compatible with any OpenAI-compatible endpoint
• ✅ Local Models — vLLM, Ollama, LM Studio support
• ✅ Code Sandboxes — JavaScript (Node.js vm) and Python (system subprocess) execution
• ✅ Document Support — PDF and DOCX text extraction (bundled, works out of the box)
• ✅ Web Search — Tavily and Z.AI integration for internet search
• ✅ Telegram Parsing — render t.me channels with reactions, views, auto-scroll for older posts
• ✅ Security — directory sandboxing for safe file operations
• ✅ Cross-platform — Windows, macOS, Linux with proper shell commands

UI/UX Features

• ✅ Modern Interface — React + Tauri with smooth auto-scroll and streaming
• ✅ Message Editing — edit and resend messages with history truncation
• ✅ Session Persistence — sessions survive app restart (SQLite backed)
• ✅ Session Management — pin important sessions, search through chat history
• ✅ Keyboard Shortcuts — Cmd+Enter/Ctrl+Enter to send messages
• ✅ Spell Check — built-in spell checking with context menu suggestions
• ✅ Permission System — ask/default modes for tool execution control

Advanced Features

• ✅ Memory System — persistent storage of user preferences in ~/.valera/memory.md
• ✅ Token Tracking — display input/output tokens and API duration
• ✅ Optimized Streaming — requestAnimationFrame-based UI updates (60fps)
• ✅ Stop Streaming — interrupt LLM responses at any time
• ✅ Loop Detection — automatic detection of stuck tool call loops (5+ sequential same-tool calls)
• ✅ Request Timeouts — 5-minute timeout with auto-retry for LLM requests
• ✅ Session Logging — full request/response JSON logs per iteration in ~/.valera/logs/sessions/

🤔 Why ValeDesk?

Open Architecture & Full Control

ValeDesk isn't just another AI assistant — it's a framework you own. Built with TypeScript and Tauri, every component is transparent and modifiable:

• Readable codebase — well-structured, documented code you can understand
• Easy customization — add new tools, modify prompts, change UI without black boxes
• Your rules — adjust behavior, safety limits, and workflows to match your needs
• No vendor lock-in — works with any OpenAI-compatible API (vLLM, Ollama, LM Studio)

100% Local & Private

Everything runs on your machine:

• Local inference — use Ollama, vLLM, or LM Studio for complete privacy
• No data collection — your conversations never leave your computer
• Offline capable — works without internet (except web search tools)
• Sandboxed execution — secure JavaScript sandbox and file operation restrictions

Experiment & Iterate

Perfect for developers, researchers, and AI enthusiasts:

• Test local models — compare Qwen, Llama, DeepSeek, and others
• Debug API calls — full request/response logs for every interaction
• Prototype tools — add custom functions in minutes
• Monitor performance — track tokens, timing, and resource usage

Real Use Cases

# Run Ollama locally (free, 100% private)
ollama serve
# Configure ValeDesk: http://localhost:11434/v1

# Or use vLLM for faster inference
vllm serve Qwen/Qwen2.5-14B-Instruct --port 8000
# Configure ValeDesk: http://localhost:8000/v1

TL;DR: ValeDesk gives you the power of ChatGPT/Claude with the freedom of open source and privacy of local execution.

🚀 Quick Start

Prerequisites

• Rust 1.74+ (install)
• Node.js 20+
• Python 3 (for execute_python tool)

Development (macOS/Linux)

# Clone and enter
git clone https://github.com/vakovalskii/ValeDesk.git
cd ValeDesk

# Install dependencies
npm install

# Run in development mode
make dev

Tests

npm run test

Build Standalone App

# Build DMG (macOS)
make bundle

# Output: ValeDesk-0.0.8.dmg

Manual Build Steps

# 1. Build sidecar binary
npm run build:sidecar

# 2. Build Tauri app
cd src-tauri && cargo build --release

# 3. Create DMG
hdiutil create -volname "ValeDesk" \
  -srcfolder src-tauri/target/release/bundle/macos/ValeDesk.app \
  -ov -format UDZO ValeDesk-0.0.8.dmg

Windows (coming soon)

Windows build requires cross-compilation setup. Check .github/workflows/ for CI builds.

Configuration

1. Click Settings (⚙️) in the app
2. Configure your API:
   • API Key — your key (or dummy-key for local models)
   • Base URL — API endpoint (must include /v1)
   • Model Name — model identifier
   • Temperature — 0.0-2.0 (default: 0.3)
3. Click Save Settings

Voice Input (Speech-to-Text)

• Start a local transcription server:
  • chmod +x scripts/setup_voice_server.sh
  • ./scripts/setup_voice_server.sh (Docker by default; falls back to uvx or venv)
  • Optional vars: PORT, MODEL, DEVICE, COMPUTE_TYPE, DOCKER_TAG
  • Docker resources: DOCKER_MEMORY (e.g. 6g), DOCKER_CPUS (e.g. 6), DOCKER_SHM_SIZE (e.g. 2g)
  • Example: DOCKER_MEMORY=6g DOCKER_CPUS=6 DOCKER_SHM_SIZE=2g ./scripts/setup_voice_server.sh
• In Settings → Voice set Voice Base URL (e.g. http://localhost:8000/v1), Model, optional API Key/Language
• Wait for Connected, then use the mic button in the prompt input

Example Configurations

Local vLLM:

{
  "apiKey": "dummy-key",
  "baseUrl": "http://localhost:8000/v1",
  "model": "qwen3-30b-a3b-instruct-2507"
}

OpenAI:

{
  "apiKey": "sk-...",
  "baseUrl": "https://api.openai.com/v1",
  "model": "gpt-4"
}

🎯 Skills Marketplace

Browse and install verified skills for ValeDesk: Skills Marketplace

🛠️ Available Tools

All tools follow snake_case naming convention (verb_noun pattern):

File Operations

Code Execution

Web Tools

Task Management

Scheduler

Features:

• One-time reminders — "remind me in 30 minutes"
• Recurring tasks — every minute, hour, day, week, month
• Auto-execution — tasks with prompts automatically start new chat sessions
• Native notifications — macOS system notifications
• Default model — set preferred model for scheduled tasks

Memory

Security: All file operations are sandboxed to the workspace folder only.

📦 Building

Windows

# Build executable and installer
npm run dist:win

# Output: dist/ValeDesk Setup 0.0.8.exe

macOS

# Build DMG (ARM64)
npm run dist:mac-arm64

# Build DMG (Intel x64)
npm run dist:mac-x64

Linux

# Build AppImage
npm run dist:linux

🔐 Data Storage

Application Data

• Windows: C:\Users\YourName\AppData\Roaming\ValeDesk\
• macOS: ~/Library/Application Support/ValeDesk/
• Linux: ~/.config/ValeDesk/

Files:

• sessions.db — SQLite database with chat history, todos, scheduled tasks, and settings
• api-settings.json — API configuration
• skills-settings.json — Skills marketplace configuration

Global Data

• ~/.valera/memory.md — persistent memory storage
• ~/.valera/logs/sessions/{session-id}/ — per-session API logs:
  • turn-001-request.json — full request (model, messages, tools, temperature)
  • turn-001-response.json — full response (usage, content, tool_calls)

🛠️ Contributing

See CURSOR.md for development guidelines and project architecture.

⭐ Star History

📄 License

ValeDesk Community License — free for individuals and companies with reven