The Problem

AI-generated code works, but rarely passes code review:

Sound familiar? You spend more time fixing AI code than writing it yourself.

Corbat MCP solves this by injecting your team's coding standards before the AI generates code — not after.

Quick Start

1. Add to your MCP config:

{
  "mcpServers": {
    "corbat": {
      "command": "npx",
      "args": ["-y", "@corbat-tech/coding-standards-mcp"]
    }
  }
}

2. Config file location:

Complete setup guide for all 25+ tools

3. Done! Corbat auto-detects your stack and applies the right standards.

Zero overhead. Corbat runs locally and adds ~50ms to detect your stack. After that, it's just context for the AI.

Fully customizable. Don't like a rule? Override it in .corbat.json. Jump to customization →

See It In Action

You: "Create a user registration service"

Without Corbat, you get:

class UserService {
  private users: Map<string, User> = new Map();

  register(name: string, email: string, password: string) {
    if (!name || !email || !password) throw new Error('Invalid input');
    const user = { id: crypto.randomUUID(), name, email, password };
    this.users.set(user.id, user);
    return user;
  }
}
// ✗ Hardcoded storage  ✗ Password in plain text  ✗ Generic error  ✗ No tests

With Corbat, you get:

// Domain
interface UserRepository {
  save(user: User): Promise<void>;
  findByEmail(email: string): Promise<User | null>;
}

class EmailAlreadyExistsError extends Error {
  constructor(email: string) {
    super(`Email already registered: ${email}`);
  }
}

// Application
class RegisterUserUseCase {
  constructor(
    private readonly repository: UserRepository,
    private readonly hasher: PasswordHasher
  ) {}

  async execute(input: RegisterUserInput): Promise<User> {
    const existing = await this.repository.findByEmail(input.email);
    if (existing) throw new EmailAlreadyExistsError(input.email);

    const user = User.create({
      ...input,
      password: await this.hasher.hash(input.password)
    });
    await this.repository.save(user);
    return user;
  }
}
// ✓ Repository interface  ✓ Password hashing  ✓ Custom error  ✓ Testable

// Test included
describe('RegisterUserUseCase', () => {
  const repository = { save: vi.fn(), findByEmail: vi.fn() };
  const hasher = { hash: vi.fn() };
  const useCase = new RegisterUserUseCase(repository, hasher);

  beforeEach(() => vi.clearAllMocks());

  it('should hash password before saving', async () => {
    repository.findByEmail.mockResolvedValue(null);
    hasher.hash.mockResolvedValue('hashed_password');

    await useCase.execute({ name: 'John', email: 'john@test.com', password: 'secret' });

    expect(hasher.hash).toHaveBeenCalledWith('secret');
    expect(repository.save).toHaveBeenCalledWith(
      expect.objectContaining({ password: 'hashed_password' })
    );
  });

  it('should reject duplicate emails', async () => {
    repository.findByEmail.mockResolvedValue({ id: '1', email: 'john@test.com' });

    await expect(
      useCase.execute({ name: 'John', email: 'john@test.com', password: 'secret' })
    ).rejects.toThrow(EmailAlreadyExistsError);
  });
});

This is what "passes code review on the first try" looks like.

What Corbat Enforces

Corbat injects these guardrails before code generation:

Code Quality

Architecture

Error Handling

Security (verified against OWASP Top 10)

Benchmark Results v3.0

We evaluated Corbat across 15 real-world scenarios in 6 languages.

The Key Insight

Corbat generates focused, production-ready code — not verbose boilerplate:

This isn't "less code for less code's sake" — it's the right abstractions without over-engineering.

Value Metrics

When we measure what actually matters for production code:

📊 Detailed value analysis

Security: Zero Vulnerabilities Detected

Every scenario was analyzed using pattern detection for OWASP Top 10 vulnerabilities:

• ✓ No SQL/NoSQL injection patterns
• ✓ No XSS vulnerabilities
• ✓ No hardcoded credentials
• ✓ Input validation at all boundaries
• ✓ Proper error messages (no stack traces to users)

Languages & Patterns Tested

📖 Full benchmark methodology · Value analysis

Built-in Profiles

Corbat auto-detects your stack and applies the right standards:

Auto-detection: Corbat reads pom.xml, package.json, go.mod, Cargo.toml, etc.

When to Use Corbat

When Corbat Might Not Be Needed

• Quick prototypes where quality doesn't matter
• One-off scripts you'll throw away
• Learning projects where you want to make mistakes

Customize

Option 1: Interactive Setup

npx corbat-init

Detects your stack and generates a .corbat.json with sensible defaults.

Option 2: Manual Configuration

Create .corbat.json in your project root:

{
  "profile": "java-spring-backend",
  "architecture": {
    "pattern": "hexagonal",
    "layers": ["domain", "application", "infrastructure", "api"]
  },
  "quality": {
    "maxMethodLines": 20,
    "maxClassLines": 200,
    "minCoverage": 80
  },
  "rules": {
    "always": [
      "Use records for DTOs",
      "Prefer Optional over null"
    ],
    "never": [
      "Use field injection",
      "Catch generic Exception"
    ]
  }
}

Option 3: Use a Template

Browse 14 ready-to-use templates for Java, Python, Node.js, React, Go, Rust, and more.

How It Works

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│ Your Prompt │────▶│ Corbat MCP  │────▶│ AI + Rules  │
└─────────────┘     └──────┬──────┘     └─────────────┘
                          │
           ┌──────────────┼──────────────┐
           ▼              ▼              ▼
    ┌────────────┐ ┌────────────┐ ┌────────────┐
    │ 1. Detect  │ │ 2. Load    │ │ 3. Inject  │
    │   Stack    │ │  Profile   │ │ Guardrails │
    └────────────┘ └────────────┘ └────────────┘
     pom.xml        hexagonal      max 20 lines
     package.json   + DDD          + interfaces
     go.mod         + SOLID        + custom errors

Corbat doesn't modify AI output — it ensures the AI knows your standards before generating.

Important: Corbat provides context and guidelines to the AI. The actual code quality depends on how well the AI model follows these guidelines. In our testing, models like Claude and GPT-4 consistently respect these guardrails.

Documentation

{ "mcpServers": { "corbat": { "command": "npx", "args": ["-y", "@corbat-tech/coding-standards-mcp"] }}}