<div align="center">
CORBAT MCP
Policy and quality engine for AI coding agents
Shared standards, context, and verification gates for agent-assisted software delivery.
    
---
     
Designed for MCP-compatible coding tools and multi-agent workflows. See compatibility notes.
⚡ Try it in 30 seconds — just add the config below and start coding.
</div>
---
The Problem
AI-generated code often works locally but misses team-specific review expectations:
| Without Corbat | With Corbat | |----------------|-------------| | Methods with 50+ lines | Max 20 lines per method | | No dependency injection | Proper DI with interfaces | | throw new Error('failed') | Custom exceptions with context | | Missing or minimal tests | Tests included, TDD approach | | God classes, mixed concerns | SOLID principles, clean layers | | Works locally | Reviewable implementation with explicit checks |
Corbat helps by giving agents your team's standards before implementation and by providing validation tools before review or handoff.
---
Quick Start
1. Add to your MCP config:
{
"mcpServers": {
"corbat": {
"command": "npx",
"args": ["-y", "@corbat-tech/coding-standards-mcp"]
}
}
}2. Config file location:
| Tool | Location | |------|----------| | Cursor | .cursor/mcp.json | | VS Code | .vscode/mcp.json | | Windsurf | ~/.codeium/windsurf/mcp_config.json | | JetBrains | Settings → AI Assistant → MCP | | Claude Desktop | ~/.config/Claude/claude_desktop_config.json | | Claude Code | claude mcp add corbat -- npx -y @corbat-tech/coding-standards-mcp |
Complete setup guide for common MCP-compatible tools
3. Done. Corbat can auto-detect your stack and return matching 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 testsWith 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 the kind of structure Corbat asks agents to consider before handoff.
---
What Corbat Provides
Corbat provides these guardrails before generation and validation checks during iteration:
Code Quality
| Rule | Why It Matters | |------|----------------| | Max 20 lines per method | Readable, testable, single-purpose functions | | Max 200 lines per class | Single Responsibility Principle | | Meaningful names | No data, info, temp, x | | No magic numbers | Constants with descriptive names |
Architecture
| Rule | Why It Matters | |------|----------------| | Interfaces for dependencies | Testable code, easy mocking | | Layer separation | Domain logic isolated from infrastructure | | Hexagonal/Clean patterns | Framework-agnostic business rules |
Error Handling
| Rule | Why It Matters | |------|----------------| | Custom exceptions | UserNotFoundError vs Error('not found') | | Error context | Include IDs, values, state in errors | | No empty catches | Every error handled or propagated |
Security-Oriented Checks
| Rule | Why It Matters | |------|----------------| | Input validation | Reject bad data at boundaries | | No hardcoded secrets | Environment variables only | | Parameterized queries | Prevent SQL injection | | Output encoding | Prevent XSS |
---
Benchmark Results v3.0
We evaluated Corbat across 15 scenarios in 6 languages. The primary benchmark report is mixed and should be read honestly: Corbat won 1/15 scenarios by the original aggregate score, while an alternative value analysis found stronger results for code compactness and maintainability.
What The Data Supports
The strongest observed signal is that Corbat-guided outputs are often smaller and more focused:
| Scenario | With Corbat | Without Corbat | What This Means | |----------|:-----------:|:--------------:|-----------------| | Kotlin Coroutines | 236 lines | 1,923 lines | Same functionality, 8x less to maintain | | Java Hexagonal | 623 lines | 2,740 lines | Clean architecture without the bloat | | Go Clean Arch | 459 lines | 2,012 lines | Idiomatic Go, not Java-in-Go | | TypeScript NestJS | 395 lines | 1,554 lines | Right patterns, right size |
This is not enough to claim universal quality improvement. It is evidence that standards context can reduce over-generation in some workflows.
Value Metrics
The value report reweights the same dataset toward efficiency and maintainability:
| Metric | Result | What It Means | |--------|:------:|---------------| | Code Reduction | 67% | Less to maintain, review, and debug | | Security checks | 100% | No issues detected by benchmark pattern checks | | Maintainability | 93% win | Easier to understand and modify | | Architecture Efficiency | 87% win | Better patterns per line of code | | Cognitive Load | -59% | Faster onboarding for new developers |
Security: No Benchmark Pattern Findings
Every scenario was analyzed using pattern checks inspired by OWASP Top 10 categories. This is not a replacement for SAST, DAST, dependency scanning, manual review, or threat modeling.
- No SQL/NoSQL injection patterns detected
- No XSS patterns detected
- No hardcoded credentials detected
- Input validation patterns present at boundaries
- Error messages did not expose stack traces in the benchmark samples
Languages & Patterns Tested
| Language | Scenarios | Patterns | |:--------:|:---------:|:---------| | ☕ Java | 5 | Spring Boot, DDD Aggregates, Hexagonal, Kafka Events, Saga | | 📘 TypeScript | 4 | Express REST, NestJS Clean, React Components, Next.js Full-Stack | | 🐍 Python | 2 | FastAPI CRUD, Repository Pattern | | 🐹 Go | 2 | HTTP Handlers, Clean Architecture | | 🦀 Rust | 1 | Axum with Repository Trait | | 🟣 Kotlin | 1 | Coroutines + Strategy Pattern |
📖 Full benchmark methodology · Value analysis
---
Built-in Profiles
Corbat auto-detects your stack and applies the right standards:
| Profile | Stack | What You Get | |---------|-------|--------------| | java-spring-backend | Java 21 + Spring Boot 3 | Hexagonal + DDD, TDD with 80%+ coverage | | kotlin-spring | Kotlin + Spring Boot 3 | Coroutines, Kotest + MockK | | nodejs | Node.js + TypeScript | Clean Architecture, Vitest | | nextjs | Next.js 14+ | App Router patterns, Server Components | | react | React 18+ | Hooks, Testing Library, accessible components | | vue | Vue 3.5+ | Composition API, Vitest | | angular | Angular 19+ | Standalone components, Jest | | python | Python + FastAPI | Async patterns, pytest | | go | Go 1.22+ | Idiomatic Go, table-driven tests | | rust | Rust + Axum | Ownership patterns, proptest | | csharp-dotnet | C# 12 + ASP.NET Core 8 | Clean + CQRS, xUnit | | flutter | Dart 3 + Flutter | BLoC/Riverpod, widget tests |
Auto-detection: Corbat reads pom.xml, package.json, go.mod, Cargo.toml, etc.
---
When to Use Corbat
| Use Case | Why Corbat Helps | |----------|------------------| | Starting a new project | Correct architecture from day one | | Multi-agent delivery | Planner, implementation, review, and security agents share the same policy context | | Teams with mixed experience | Standards become explicit and repeatable | | Strict code review standards | Agents can validate against the review bar before handoff | | Regulated industries | Consistent security and documentation | | Legacy modernization | New code follows modern patterns |
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-initDetects 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 errorsCorbat does not modify AI output. It supplies context, standards, profiles, and validation feedback so agents can align with your review expectations.
Important: Actual code quality depends on the model, host tool, prompt, repository context, tests, and human review. Treat Corbat as a policy and verification layer, not as a guarantee of production readiness.
---
Documentation
| Resource | Description | |----------|-------------| | Setup Guide | Installation notes for common MCP-compatible tools | | Templates | Ready-to-use .corbat.json configurations | | Compatibility | Full list of supported AI tools | | Benchmark Analysis | Detailed results from 15 scenarios | | Multi-Agent Architecture | How Corbat fits planner/reviewer/security/release agent workflows | | Security Model | Local execution model, threat model, and reporting | | API Reference | Tools, prompts, and configuration options |
---
<div align="center">
Make agent output easier to review.
Add to your MCP config and you're done:
{ "mcpServers": { "corbat": { "command": "npx", "args": ["-y", "@corbat-tech/coding-standards-mcp"] }}}Use Corbat as shared standards context plus a local quality gate.
---
Developed by corbat-tech
</div>
