Gorstak's IDE — a free, fully local AI coding assistant for Windows, built on .NET 4.8.
GIDE runs entirely on your machine using llama.cpp and Qwen3 models. No cloud, no API keys, no usage limits. Point it at a project folder, describe what you want, and it writes code directly into your files.
- 100% local — no API keys, no internet required after setup, no data leaves your machine.
- Auto-downloads llama.cpp — always fetches the latest release from GitHub automatically.
- Qwen3 model support — choose from 4B, 8B, 14B, or 30B A3B models depending on your hardware.
- Hardware-aware — detects RAM, CPU cores, and CUDA GPU to pick the optimal model and build.
- Auto port selection — finds a free port automatically, no manual configuration needed.
- Writes code directly — model outputs are parsed and executed as file writes, not just displayed.
- Project-aware — reads all source and documentation files in your project for full context.
- Conversation memory — full chat history is sent to the model and persisted to disk.
- Bottom-docked prompt — input stays at the bottom, giving AI responses maximum vertical space.
- Chrome-dark theme — clean dark UI with Mica backdrop (Win11) and region-clipped rounded buttons.
- Project browser — open a folder and GIDE will scan and provide context about your codebase.
- Context menu integration — right-click any folder in Explorer and choose "Run GIDE here".
- Windows 10 or later (64-bit)
- .NET Framework 4.8 (pre-installed on Windows 10/11)
- Free disk space for models:
- Qwen3 4B: ~2.5 GB
- Qwen3 8B: ~5 GB
- Qwen3 14B: ~9 GB (recommended)
- Qwen3 30B A3B: ~18.6 GB
- RAM: 4 GB minimum, 16 GB recommended for 14B model
- Internet connection on first launch (to download llama.cpp and chosen model)
Download GIDE-Setup-0.8.0.exe from the releases page and run it.
The installer:
- Installs GIDE to
Program Files\GIDE - Downloads and installs the Visual C++ Redistributable if missing (required by llama.cpp)
- Adds "Run GIDE here" to the folder right-click context menu
- Creates Start Menu and optional desktop shortcuts
git clone https://github.com/CroatiaSecurity/GIDE
cd GIDE
build.cmd
The compiled binary lands at dist\GIDE.exe.
Requires the .NET Framework 4.8 SDK / csc.exe (included with Visual Studio or the .NET 4.8 Developer Pack).
On first launch GIDE will automatically:
- Query GitHub for the latest llama.cpp release and download it (~30 MB)
- Prompt you to select and download a Qwen3 model
- Start the local inference server and begin responding
This is a one-time setup. After that, GIDE starts instantly with no internet needed.
Use the dropdown in the top-right corner to select a model. The recommended model for your hardware is pre-selected.
| Model | Size | RAM Required |
|---|---|---|
| Qwen3 4B | ~2.5 GB | 4 GB |
| Qwen3 8B | ~5 GB | 8 GB |
| Qwen3 14B (recommended) | ~9 GB | 16 GB |
| Qwen3 30B A3B | ~18.6 GB | 32 GB / 12 GB VRAM |
Click Download to download a model. Progress is shown in the status bar.
- GIDE starts a local
llama-serverprocess (llama.cpp) with your selected model. - Your messages are sent to the local server via the OpenAI-compatible
/v1/chat/completionsAPI. - The model generates a response locally (timeout: 30 minutes for large contexts).
- Tool calls (
<<<TOOL:WRITE>>>,<<<TOOL:RUN>>>) are parsed and executed automatically. - Markdown code blocks with file paths are auto-converted to file writes.
- Qwen3
<think>blocks are automatically stripped from the output.
When you open a folder, GIDE automatically:
- Scans all files recursively (skipping
bin,obj,node_modules,.git, etc.) - Reads source and documentation files (blacklist-based: skips binaries, media, archives)
- Includes file contents in the system prompt so the model understands your codebase
- Sends conversation history for multi-turn context
- All inference runs locally on your hardware.
- No telemetry, no analytics, no remote logging.
- Models are stored in
%USERPROFILE%\.gide\models\. - llama.cpp binaries are stored in
%USERPROFILE%\.gide\bin\. - Chat history is saved to
.gide_history.jsonin the project folder.
Requires Inno Setup 6 installed to C:\Program Files (x86)\Inno Setup 6\.
build.cmd
Output: releases\0.8.0\GIDE-Setup-0.8.0.exe
- Context Window Fix: Stopped injecting entire project file contents into the prompt, preventing context overflow and improving speed. Added support for
<<<TOOL:READ>>>instead. - Port Assignment Fix: Resolved port exhaustion by dynamically letting the OS assign ephemeral ports.
- Markdown Inference Fix: Fixed ambiguous file overwrites when inferring tool commands from markdown.
- Tool execution in GUI: Model responses with
<<<TOOL:WRITE>>>markers or markdown code blocks are now parsed and executed automatically (file creation/editing). - Full project context: System prompt includes actual file contents from the project (not just filenames), with blacklist-based filtering to skip binaries.
- Conversation memory: Full chat history is sent to the model for multi-turn context, with smart trimming to fit the context window.
- History persistence: Chat is saved to
.gide_history.jsonin the project folder, reloaded on next session. - 30-minute timeout: HTTP timeout increased to 30 minutes for large context generation on CPU.
- Better error reporting: 400/timeout errors from llama-server now show the server's response body for debugging.
- Recursive file scanning: Reads all source files in subdirectories (skipping build/binary folders).
- VC++ Redistributable: Installer now downloads and installs the Visual C++ Runtime automatically via curl.
- New layout: Prompt/input moved to the bottom of the window, giving AI responses the full vertical space above.
- Chrome-dark theme: Switched from purple/creamy-black to a flat dark palette matching GBrowser.
- Mica backdrop: DWM Mica effect with subtle gradient glow for visual depth (Win11 22H2+).
- Region-clipped buttons: Rounded buttons use Region clipping, eliminating corner artifacts.
- Download progress: Real-time percentage, MB downloaded, and speed shown in status bar text.
- Fixed model URLs: All download URLs now point to official Qwen GGUF repos (4B, 8B, 14B, 30B-A3B).
- Tool execution: GIDE can create, edit, and delete files via
<<<TOOL:WRITE>>>/<<<TOOL:RUN>>>markers in model output. - Markdown-to-tool conversion: Responses with code blocks are auto-converted to file-write operations when a path can be inferred.
- Model download progress: Status bar shows percentage + speed during model downloads.
- Settings form: Added Tools > Model Manager dialog for managing downloaded models.
- History manager: Chat history is persisted across sessions.
- Improved error reporting: Model engine failures show detailed diagnostics inline in chat.
- Context menu installer: "Run GIDE here" added to Windows Explorer folder context menu.
See LICENSE.