Skip to content

README.md

Latest commit

 

History

History
338 lines (248 loc) · 8.48 KB

README.md

File metadata and controls

338 lines (248 loc) · 8.48 KB

Multi-Model AI Toolkit

Leverage OpenAI, Anthropic, and Google Gemini together for superior code quality.

Part of Claude Code Sidekick

Features

  • 🔍 Consensus Code Review - Run code through multiple models, find issues all agree on
  • 🎯 Intelligent Routing - Automatically route tasks to the optimal model
  • 🔎 Semantic Search - Index your codebase and search by meaning, not keywords
  • 💰 Cost Optimization - Use cheap models for simple tasks, reserve power for complex ones
  • 🔧 Flexible Configuration - Works with just one API key or all three

File Structure

tools/multi-model/
├── bin/
│   ├── review.js          # mm-review CLI
│   ├── index-codebase.js  # mm-index CLI
│   └── search.js          # mm-search CLI
├── lib/
│   ├── clients.js         # Unified API for all 3 providers
│   ├── code-review.js     # Consensus review system
│   ├── model-router.js    # Intelligent task routing
│   └── embeddings.js      # Codebase indexing & search
├── examples/
│   ├── pre-commit-hook.js # Git hook integration
│   └── agent-integration.js # Claude Code workflows
├── index.js               # Main exports
├── package.json
└── README.md

Quick Start

# Install dependencies
cd tools/multi-model
pnpm install

# Set up your API keys in .env (root of claude-code-sidekick)
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=...

Configuration

Single-Model vs Multi-Model Mode

The toolkit automatically adapts based on your available API keys:

API Keys Available Mode Behavior
All three Multi-model Consensus review across all models
Two keys Multi-model Consensus with 2 models
One key Single-model Direct review with one model

To explicitly disable multi-model mode (even if you have multiple keys):

# In your .env file
USE_MULTI_MODEL=FALSE

This is useful for:

  • Reducing API costs
  • Faster reviews (single API call)
  • Testing with a specific model

Minimum Requirements

You only need one API key to use this toolkit:

# Option 1: OpenAI only
OPENAI_API_KEY=sk-...

# Option 2: Anthropic/Claude only
ANTHROPIC_API_KEY=sk-ant-...

# Option 3: Gemini only
GEMINI_API_KEY=...

The toolkit will automatically use the best available model for your configuration.

CLI Tools

Multi-Model Code Review

# Quick review (Gemini Flash + GPT-4o-mini)
pnpm review -- src/app.js --quick

# Deep review (GPT-4o + Claude Sonnet + Gemini Pro)
pnpm review -- src/app.js --deep

# Custom models
pnpm review -- src/app.js --models "gpt-4o,claude-3-5-sonnet-20241022"

# Output as JSON (for CI integration)
pnpm review -- src/app.js --json

Codebase Indexing

# Index current directory
pnpm index

# Index specific directory
pnpm index -- /path/to/project

# With options
pnpm index -- . --verbose --extensions ".ts,.tsx"

Semantic Search

# Search your codebase
pnpm search -- "authentication middleware"

# More results, lower threshold
pnpm search -- "error handling" --top 20 --similarity 0.4

# JSON output
pnpm search -- "database connection" --json

Programmatic Usage

Code Review

import {
  reviewCode,
  quickReview,
  deepReview,
} from '@claude-sidekick/multi-model';

// Basic review (3 models, consensus required)
const results = await reviewCode(code, {
  filename: 'auth.js',
  consensusThreshold: 2, // 2+ models must agree
});

console.log(results.confirmedIssues); // Issues multiple models found
console.log(results.possibleIssues); // Issues only one model found

// Quick review for CI
const quick = await quickReview(code, 'app.js');

// Deep review for important changes
const deep = await deepReview(code, 'payment.js');

Intelligent Model Routing

import { createRouter, ModelRouter } from '@claude-sidekick/multi-model';

// Use a preset profile
const router = createRouter('balanced'); // 'cost', 'speed', 'quality', 'balanced'

// Tasks are automatically routed to optimal models
const result = await router.route(
  'Explain the architecture of this codebase...'
);
console.log(result.model); // e.g., 'claude-3-5-sonnet-20241022'
console.log(result.taskType); // e.g., 'architecture'
console.log(result.routing); // 'Claude excels at complex reasoning...'

// Estimate costs
const cost = router.estimateCost('documentation', 5000, 2000);
console.log(cost.formatted); // '$0.0215'

Semantic Code Search

import {
  indexCodebase,
  searchCodebase,
  findSimilarCode,
} from '@claude-sidekick/multi-model';

// Index a codebase (do this once, or on changes)
await indexCodebase('/path/to/project', {
  extensions: ['.js', '.ts', '.tsx'],
  verbose: true,
});

// Search for code by meaning
const results = await searchCodebase('user authentication logic', {
  indexPath: '/path/to/project/.code-index/index.json',
  topK: 10,
  minSimilarity: 0.5,
});

// Find duplicate/similar code
const similar = await findSimilarCode(myCode, {
  indexPath: '/path/to/project/.code-index/index.json',
  threshold: 0.85,
});

Model Selection Guide

Task Best Model Why
Long file analysis Claude Sonnet 200k context window
Code generation GPT-4o Mature function calling
Quick refactors GPT-4o-mini Fast, accurate
Documentation Claude Sonnet Nuanced, thorough
Image analysis Gemini Pro Native multimodal
Security review Claude Sonnet Careful reasoning
Linting/style Gemini Flash Cheapest

Cost Comparison (per 1M tokens)

Model Input Output
Gemini Flash $0.075 $0.30
GPT-4o-mini $0.15 $0.60
Claude Haiku $0.25 $1.25
Claude Sonnet $3.00 $15.00
Gemini Pro $3.50 $10.50
GPT-4o $5.00 $15.00

CI/CD Integration

# GitHub Actions example
- name: Multi-Model Code Review
  run: |
    cd tools/multi-model
    pnpm review -- ${{ github.event.pull_request.changed_files }} --json > review.json

- name: Fail on Critical Issues
  run: |
    if jq -e '.confirmedIssues[] | select(.severity == "critical")' review.json; then
      echo "Critical issues found!"
      exit 1
    fi

Environment Variables

# Required (at least one)
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=...  # or GEMENI_API_KEY (typo-tolerant)

# Optional
DEBUG=true          # Enable debug logging

Examples

Pre-Commit Hook

Block commits that have critical/high severity issues detected by multiple models.

📄 examples/pre-commit-hook.js

// Install in .husky/pre-commit or run directly
import { quickReview } from '../lib/code-review.js';

// Reviews staged files, blocks on critical issues
const results = await quickReview(code, filename);
if (results.confirmedIssues.some((i) => i.severity === 'critical')) {
  process.exit(1); // Block commit
}

Agent Integration

Smart assistant that routes tasks to optimal models and uses semantic search for context.

📄 examples/agent-integration.js

import { createRouter, searchCodebase } from '../index.js';

const router = createRouter('balanced');

// Search for relevant code context
const context = await searchCodebase('authentication', { indexPath });

// Route to optimal model based on task type
const result = await router.route(userRequest, { context });
// → Automatically picks Claude for architecture, GPT-4 for code gen, etc.

Supported Models

OpenAI

  • gpt-4o - Most capable, best for complex tasks
  • gpt-4o-mini - Fast and cheap, good for simple tasks

Anthropic

  • claude-sonnet-4-20250514 - Best reasoning, 200k context
  • claude-3-5-haiku-20241022 - Fast and cheap

Google

  • gemini-1.5-pro - Multimodal, good reasoning
  • gemini-2.0-flash - Very fast, very cheap

Related Documentation

License

Proprietary - © 2026 Dylan Burkey