DeskAI — Intelligent Local Coding Assistant

100% offline AI coding assistant powered by local LLMs running entirely on your machine. No API keys. No data sent to the cloud. Full project generation, agentic file editing, and terminal control — all on your hardware.

Features

🤖 Agent Mode — Full Project Generation

Let the AI plan and build entire projects autonomously:

• Creates files and folders directly on disk
• Runs terminal commands (npm install, build, etc.)
• Reads, edits, and verifies files
• Recovers from errors automatically
• Persistent task memory (scratchpad) survives context resets
• "Continue task" button resumes interrupted sessions

💬 Ask Mode — Instant Code Help

• Code explanation, refactoring, debugging
• Ask about selected code via right-click → DeskAI: Ask About Selection
• Attach files as context with the 📎 button

📎 File Attachment

• Attach single files or entire folders as context
• File content processed server-side — never floods the chat UI
• Large files handled gracefully with truncation and read_file fallback

🔋 Model Power Toggle

• One-click unload the model from RAM when you're done (saves 8–20 GB of memory)
• Reload instantly without restarting VS Code
• Power button in the chat header — green = loaded, dim = unloaded

🕘 Persistent Chat History

• Every conversation is automatically saved to local storage
• Reopen VS Code and restore any previous session with one click
• History panel shows session titles and timestamps
• Delete individual sessions or start fresh with Clear

🔍 Workspace RAG (Retrieval-Augmented Generation)

• Automatically injects relevant workspace context into every query — no manual file attaching
• Keyword-based retrieval scans up to 300 files across your project
• Works entirely offline: no embeddings API, no cloud
• Configurable depth (deskai.ragTopK) and can be toggled off

⚡ Local LLM Powered

• Uses llama-server under the hood (bundled)
• Supports GGUF models — import any compatible model
• Recommended models auto-downloaded on first launch

Recommended Models

Models are downloaded automatically on first use. You can also Import any .gguf file via the model picker.

Getting Started

1. Install the extension
2. Open the DeskAI panel in the secondary sidebar (right side)
3. Wait for the model to load (first launch downloads the model ~8–16 GB)
4. Switch to Agent mode for full project generation
5. Type your task and press Enter

Showcase / Demo Mode

For full project generation without confirmation dialogs:

• Enable deskai.autoApproveWrites: true — files written directly without popups
• Enable deskai.hiddenTerminal: true — commands run silently in background

Settings

| Setting | Default | Description | |---------|---------|-------------| | deskai.autoApproveWrites | false | Write files directly without confirmation dialogs | | deskai.hiddenTerminal | false | Run shell commands silently in background | | deskai.maxAgentIterations | 40 | Max tool-call iterations before pausing | | deskai.contextSize | 32768 | Model context window in tokens | | deskai.temperature | 0.3 | Generation temperature (0=deterministic) | | deskai.modelPath | | Custom path to a `.gguf` model file | | `deskai.modelsSearchPaths` | `[]` | Extra directories to scan for models | | `deskai.llamaServerPath` | | Custom path to llama-server executable | | deskai.enableRag | true | Inject relevant workspace files into every query automatically | | deskai.ragTopK | 4 | Number of workspace chunks injected per query | | deskai.loraAdapterPath | `` | Path to a LoRA adapter file (.gguf/.bin) to load with the model |

Requirements

• Windows 10/11 (primary), macOS, Linux
• RAM: 12 GB minimum (16 GB+ recommended for best models)
• Disk: 10–20 GB free for model storage
• VS Code 1.85.0+

How It Works

1. On first open, DeskAI downloads llama-server and a GGUF model
2. llama-server runs locally on port 11434 with a 32k context window
3. The chat panel communicates with the local server via HTTP
4. In Agent mode, the model calls tools (write_file, run_terminal_command, etc.) to implement tasks
5. All data stays on your machine — nothing is sent to the internet during inference

Commands

Privacy

• All inference is local — your code never leaves your machine
• The only network requests are:
  • Initial download of llama-server binary (from GitHub Releases)
  • Initial download of the selected GGUF model (from Hugging Face)
• After first setup, the extension works completely offline

License

MIT — see LICENSE