AI Guard

🌍 English | 中文 | 日本語 | Deutsch | Français | Español | Português | 한국어 | Русский | العربية | Italiano | Türkçe | हिन्दी | ภาษาไทย | Tiếng Việt

Let AIs find each other's mistakes. — Multi-model AI code assistant for VS Code with hallucination prevention.

Real dev scene: me coding in the trenches, AIs reviewing from the sidelines

Fork it, hack it, make it yours! Contributions welcome!

Overview

AI Guard uses a three-stage pipeline to generate safer, more reliable AI-assisted code:

1. Generate — A production model generates code from your prompt
2. Review — A separate review model checks the generated code for issues
3. Rule Check — Built-in and custom rules validate the output (import validation, syntax check, security patterns)

By using different models for generation and review, AI Guard catches hallucinations and errors that a single model might miss.

Features

• Multi-model pipeline — Use any OpenAI-compatible models for generation and review
• Multi-agent review — Configure multiple review agents to run in parallel for broader coverage
• ESLint integration — Static analysis via ESLint (uses workspace config or sensible defaults)
• Custom rules — Extend with your own rules via .ai-guard/rules/
• Configurable stage order — Reorder pipeline stages; optionally skip review when rule check finds errors (saves tokens)
• Diff view — Visual comparison when the review model suggests changes
• Sidebar UI — Dedicated panel showing pipeline results and status
• Configurable — Full control over models, endpoints, and pipeline behavior

Quick Start

1. Install the extension in VS Code
2. Configure your API keys in Settings → AI Guard
3. Open a file and press Ctrl+Shift+G (Cmd+Shift+G on Mac) to run the full pipeline

Commands

Command	Description
AI Guard: Generate Code	Generate code from a prompt
AI Guard: Review Selection	Review selected code
AI Guard: Run Full Pipeline	Generate + Review + Rule Check
AI Guard: Show Review Diff	Display diff between generated and reviewed code
AI Guard: Configure Models	Open model configuration panel

Configuration

Models

AI Guard supports any OpenAI-compatible API endpoint. Configure via VS Code settings:

{
  "aiGuard.productionModel.endpoint": "https://api.openai.com/v1",
  "aiGuard.productionModel.model": "gpt-4o",
  "aiGuard.productionModel.apiKey": "your-key",
  "aiGuard.reviewModel.endpoint": "https://api.openai.com/v1",
  "aiGuard.reviewModel.model": "claude-sonnet-4-20250514",
  "aiGuard.reviewModel.apiKey": "your-key"
}

Rules

Setting	Default	Description
aiGuard.rules.enableBuiltIn	true	Enable built-in rules
aiGuard.rules.enableEslint	true	Use ESLint for static analysis
aiGuard.rules.customRulesPath	.ai-guard/rules	Custom rules directory
aiGuard.rules.enableImportValidation	false	(Legacy) Check if imports exist
aiGuard.rules.enableSyntaxCheck	false	(Legacy) Validate syntax
aiGuard.rules.enableSecurityPatterns	false	(Legacy) Detect security issues

Pipeline

Setting	Default	Description
aiGuard.pipeline.autoReview	true	Auto-run review after generation
aiGuard.pipeline.autoRuleCheck	true	Auto-run rules after generation
aiGuard.pipeline.showDiffOnIssues	true	Show diff when issues found
aiGuard.pipeline.stageOrder	["ruleCheck", "review"]	Order of post-generation stages
aiGuard.pipeline.skipReviewOnError	true	Skip AI review when rule check finds errors (saves tokens)

Multi-Agent Review

You can configure multiple review agents to run in parallel. When set, this overrides the single reviewModel:

{
  "aiGuard.reviewAgents": [
    {
      "name": "Claude",
      "endpoint": "https://api.anthropic.com/v1",
      "model": "claude-sonnet-4-20250514",
      "apiKey": "your-key"
    },
    {
      "name": "DeepSeek",
      "endpoint": "https://api.deepseek.com/v1",
      "model": "deepseek-chat",
      "apiKey": "your-key"
    }
  ]
}

## Custom Rules

Create `.js` files in `.ai-guard/rules/`:

```javascript
module.exports = {
  id: 'my-rule',
  name: 'My Custom Rule',
  description: 'What this rule checks',
  async check(context) {
    const issues = [];
    // Analyze context.code, context.language, etc.
    return { issues };
  }
};

Development

npm install
npm run build    # Build extension
npm run watch    # Watch mode
npm run package  # Package as .vsix

Architecture

src/
├── extension.ts          # Entry point
├── config/               # Settings management
├── models/               # AI model providers (OpenAI-compatible)
├── pipeline/             # Three-stage pipeline engine
│   ├── generate-stage    # Stage 1: Code generation
│   ├── review-stage      # Stage 2: AI review
│   └── rule-check-stage  # Stage 3: Rule validation
├── rules/                # Hallucination detection rules
│   ├── built-in/         # Import, syntax, security checks
│   └── custom/           # Custom rule loader
├── diff/                 # Diff visualization
└── ui/                   # Sidebar, status bar, settings panel

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT