Local Ai Coder for Visual Studio Code

Local Ai Coder is a Visual Studio Code extension that embeds a fully local AI coding partner directly inside the editor. The assistant reads your project, drafts step-by-step plans, proposes shell/PowerShell commands, and prepares file changes—all of which are executed only after you explicitly approve them. No external APIs are called at runtime; inference is handled by a locally running Ollama daemon that you control.

Key capabilities

• Workspace awareness – Streams the relevant files from the open workspace into the prompt while respecting size limits and ignoring common build artifacts, with controls to fine-tune which folders are excluded.
• Local-first inference – Streams responses from an Ollama model running on your machine. Choose any Ollama model identifier in the extension settings and the assistant will download and serve it locally.
• Guarded shell access – The assistant can suggest PowerShell or shell commands. Each command is shown to you with full context and runs only after you press Approve.
• Safe file operations – File creations, edits, and deletions are surfaced as actionable items with previews and explicit confirmation prompts before changes are applied to disk.
• Interruptible workflows – Track live progress for shell commands and file operations right inside the chat view and cancel long-running plans with a single click.
• Conversation memory – Maintains a rolling, persisted conversation history so follow-up questions benefit from earlier context—even across reloads. Reset the memory at any time with the Local Ai Coder: Reset Conversation command.

Getting started

1. Install dependencies

   npm install
   

2. Compile the extension

   npm run compile
   

3. Launch the extension
   • Open this folder in Visual Studio Code.
   • Open extension.ts and press F5 to start a new Extension Development Host.
   • Run the command Local Ai Coder: Ask from the Command Palette.
4. Prepare Ollama
   • Install Ollama from ollama.com and ensure the ollama CLI is on your PATH.
   • Pull the default model by running ollama pull qwen3:4b or use Local Ai Coder: Download Model inside VS Code to fetch it automatically.

Configuration options

All settings live under the ai-code namespace. Key options include:

Use Local Ai Coder: Download Model to have the extension pull the configured Ollama model and warm the cache before you start chatting.

The Quick Actions view in the Local Ai Coder activity bar exposes shortcuts for opening the extension settings or downloading the currently selected model without crafting a prompt.

MIT © 2024 Local Dev

Offline conversion workflow

The scripts/convert_model.py helper wraps optimum so you can produce the ONNX assets transformers.js expects:

1. Ensure Python 3.10+ is available and install the conversion toolchain:

   pip install "transformers>=4.38" "optimum[onnxruntime]" onnx onnxruntime
   

2. Export the model (the example below targets Code Llama):

   npm run convert:model -- --model meta-llama/CodeLlama-7b-Instruct-hf --output "C:\models\CodeLlama7B"
   

3. Register the exported assets with Ollama using ollama create and choose a model name (for example, ollama create codellama-7b --path "C:\\models\\CodeLlama7B").
4. Update the ai-code.modelId setting to the Ollama reference you created (for example, ollama:codellama-7b) and run Local Ai Coder: Download Model to warm the cache.

The script copies config.json, tokenizer.json, and generation_config.json alongside the ONNX runtime graph in onnx/. Feel free to keep a library of converted models and swap between them by updating the ai-code.modelId setting. Ollama performs any needed quantization at runtime, so no additional steps are required here.