DEV Community

Cover image for Inside VibesPlug : The Universal 176-Skill AI Agent Swarm & MCP Infrastructure for Modern IDEs
Roedy Rustam
Roedy Rustam

Posted on Edited on Originally published at github.com

Inside VibesPlug : The Universal 176-Skill AI Agent Swarm & MCP Infrastructure for Modern IDEs

πŸ’₯ The Paradigm Shift: From Generic Prompting to Deterministic AI Swarms

AI coding assistants have taken over modern developer workflows. However, most developers still interact with AI using passive, single-prompt chat windows or basic code autocompletion. As codebases grow into microservice architectures, complex database schemas, and strictly typed ecosystems, generic AI context windows quickly degradeβ€”leading to hallucinated APIs, architectural slop, and broken state machines.

Enter vibes-plug by @roedyrustam.

vibes-plug is an open-source, cross-IDE Universal AI Agent Swarm & Model Context Protocol (MCP) plugin. Rather than treating an LLM as a simple autocomplete tool, vibes-plug converts your development environmentβ€”be it Claude Code, Cursor, Windsurf, Antigravity, or OpenAI Codexβ€”into a deterministic, multi-agent software engineering studio.

In this deep-dive article, we will unpack how vibes-plug works under the hood, dissect its 176 specialized skills, analyze its 8 specialized subagents, and examine its 8 foundational Superintelligence Engineering Engines (P1–P8).


πŸ— System Architecture Overview

vibes-plug functions as the Master Conductor over your IDE's AI agents. It intercepts low-level IDE events, structures developer intent into formal execution blueprints, and orchestrates domain-specific subagents using deterministic execution pipelines.

                      β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
                      β”‚    Developer / IDE Trigger Point       β”‚
                      β”‚ (Claude Code / Cursor / Windsurf / etc.) β”‚
                      β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                                           β”‚
                                           v
 β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
 β”‚                               VibesPlug Conductor                                β”‚
 β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
 β”‚  β€’ Skill Routing Engine        β€’ Prefix Caching (P1)      β€’ Anti-Slop Directive  β”‚
 β””β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”˜
          β”‚                                β”‚                               β”‚
          v                                v                               v
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”           β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”        β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚  Agent Topologiesβ”‚           β”‚ Superintelligence    β”‚        β”‚ Specialized Skills   β”‚
β”‚  (Subagents)     β”‚           β”‚ Engines (P1 - P8)    β”‚        β”‚ (176 Catalog Rules)  β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€           β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€        β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚ β€’ fullstack-pro  β”‚           β”‚ β€’ P1: Prompt-Cache   β”‚        β”‚ β€’ React 19 / Next 15 β”‚
β”‚ β€’ quality-guard  β”‚           β”‚ β€’ P2: In-Mem DB Sim  β”‚        β”‚ β€’ Rust Ownership    β”‚
β”‚ β€’ ai-mcp-architectβ”‚          β”‚ β€’ P3: AST Subgraph   β”‚        β”‚ β€’ Kubernetes GitOps  β”‚
β”‚ β€’ cloud-devops   β”‚           β”‚ β€’ P4: Multi-Draft PRMβ”‚        β”‚ β€’ Playwright E2E     β”‚
β”‚ β€’ sec-auditor    β”‚           β”‚ β€’ P5: SMT Verifier   β”‚        β”‚ β€’ Tailwind v4        β”‚
β”‚ β€’ mobile-expert  β”‚           β”‚ β€’ P6: Ephemeral WASM β”‚        β”‚ β€’ WebGPU / 3D        β”‚
β”‚ β€’ db-architect   β”‚           β”‚ β€’ P7: SaaT Dispatcherβ”‚        β”‚ β€’ Solana / Web3      β”‚
β”‚ β€’ systemic-refac β”‚           β”‚ β€’ P8: Schema Healer  β”‚        β”‚ β€’ ...168 more        β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜           β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜        β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
Enter fullscreen mode Exit fullscreen mode

πŸ€– The 8 Specialized Subagents (Swarm Topologies)

Instead of relying on one generic LLM prompt to write code, build DB migrations, and configure Kubernetes manifests simultaneously, vibes-plug delegates workloads to autonomous subagents operating under Fan-Out / Fan-In and Critic-Validator loops:

Subagent Scope & Core Mission Key Skill Integrations
fullstack-pro Production Fullstack Engineering React 19, Next.js 15 App Router, Tailwind v4, Hono/Bun, Prisma/Drizzle
quality-guardian Zero-Tolerance Quality & Anti-Slop TDD self-healing, Playwright E2E, adversarial security fuzzing, performance budgets
ai-mcp-architect AI & Protocol Systems MCP v1.x servers, Agent-to-Agent (A2A) interoperability, Gemini boosters
cloud-devops-sre Cloud Infrastructure & Security Kubernetes GitOps (ArgoCD/Cilium), Docker hardening, SBOM supply chain security
sec-auditor Security Auditing & Penetration OWASP Top 10, prompt injection firewalls, Zero-Trust secrets, SMT verification
mobile-crossplatform Mobile Engineering React Native (Expo Router v4), Flutter 3.x, Swift UI, Kotlin Multiplatform
db-architect Data Engineering & Migrations PostgreSQL partitioning, PGVector, Redis cluster topology, schema drift healing
systemic-refactorer Codebase Modernization & Clean Architecture AST rewriting, dead-code pruning, dependency graph decoupling

🧠 The 8 Superintelligence Engineering Engines (P1 – P8)

What truly separates vibes-plug from passive rule collections is its CLI-driven Engine Substrate. These engines inject deterministic compiler-grade guarantees into the LLM logic:

⚑ P1: Prompt-Cache Compiler (vibes compile-cache)

Compiles prefix instructions into a deterministic .cache-prefix.json (>1024 tokens). This locks in KV-Cache reuse on LLM APIs (like Anthropic & OpenAI), cutting API latency/TTFT by up to 10x and reducing token costs by 75–90%.

πŸ§ͺ P2: In-Memory DB Sandbox (vibes db:eval)

Runs an in-memory SQLite DDL/migration simulation engine to audit RLS (Row Level Security) policies, detect missing FK indexes, and prevent destructive DROP TABLE or DROP COLUMN commands before hitting staging environments.

πŸ•Έ P3: AST Subgraph Engine (vibes ast)

Parses code syntax into Abstract Syntax Trees to provide structural awareness across large codebases, avoiding naive regex matching during multi-file refactoring.

βš–οΈ P4: Speculative Multi-Draft Arbitrator (vibes arbitrate)

Implements a Process Reward Model (PRM). It evaluates competing code drafts generated by different subagents, scoring them on Type Safety, Anti-Slop Compliance, Error Resilience, and Algorithmic Complexity.

# Evaluate two competing draft solutions deterministically
vibes arbitrate draftA.ts draftB.ts
Enter fullscreen mode Exit fullscreen mode

πŸ” P5: Invariant Verifier / SMT Solver (vibes verify-invariants)

Generates SMT-LIB 2.0 formal logic specifications for theorem provers like Z3 and CVC5. It mathematically proves state machine transitions (e.g., billing, auth flows, state management) and guarantees zero unreachable states or deadlocks.

πŸ›‘ P6: Ephemeral WASM/VM Sandbox (vibes sandbox)

Executes untrusted or AI-generated scripts in an isolated, ultra-low latency (<2ms) WebAssembly sandbox with execution wall-clock timeouts and strict network isolation.

πŸš€ P7: Universal SaaT Dispatcher (vibes skill exec)

Functions as the Skills-as-a-Tool (SaaT) execution engine. It compresses and maps 176+ declarative skills and 1,000+ code blueprints into executable CLI sub-tools and context micro-slices (.skill-slices.json).

🩹 P8: Schema Drift Healer (vibes drift:heal)

Detects structural discrepancies between database DDL schemas (PostgreSQL) and application layer definitions (TypeScript/Zod/OpenAPI), automatically synthesizing a zero-drift healing patch.


πŸ›  Quickstart: How to Install and Use VibesPlug

Getting started with vibes-plug requires minimal setup:

1. Installation via NPM

Install vibes-plug globally or locally inside your workspace:

npm install -g vibes-plug
# or npx vibes-plug init
Enter fullscreen mode Exit fullscreen mode

2. Initializing Your IDE Workspace

To inject skills and agent definitions directly into your favorite editor (Cursor, Claude Code, Antigravity, etc.):

# Initialize interactive setup inside your project directory
vibes init

# Check available engine status and catalog indexing
vibes docs markdown
Enter fullscreen mode Exit fullscreen mode

3. Creating Custom Specialized Agents

You can spin up custom, tailored subagents for any project on the fly:

vibes create-agent --name "finance-auditor" --skills "rust-expert,postgresql-expert,sec-auditor"
Enter fullscreen mode Exit fullscreen mode

πŸ’‘ Practical Example: Resolving Database Schema Drift

Here is how vibes-plug P8 engine handles database schema drift autonomously when interacting with your project's code:

# Run Schema Drift Healer against PostgreSQL DDL and Zod/TypeScript schema
vibes drift:heal ./prisma/schema.prisma ./src/types/schema.ts
Enter fullscreen mode Exit fullscreen mode

Output:

{
  "status": "DRIFT_DETECTED",
  "discrepancies": [
    {
      "table": "users",
      "column": "is_mfa_enabled",
      "issue": "Present in DDL (BOOLEAN NOT NULL DEFAULT FALSE) but missing in TypeScript Zod schema."
    }
  ],
  "patch": "Added z.boolean().default(false) to UserSchema in ./src/types/schema.ts",
  "applied": true
}
Enter fullscreen mode Exit fullscreen mode

🎯 Final Thoughts

vibes-plug represents a fundamental shift in how we approach AI-assisted software engineering. By combining formal verification (SMT), AST manipulation, deterministic prefix caching, and an extensive 176-skill library, it moves developers away from unpredictable AI prompt outputs toward deterministic, production-grade software engineering swarms.

If you are building complex TypeScript, Rust, Go, Python, or Kubernetes applications and want to supercharge your IDE's AI capabilities without sacrificing code quality:

⭐ Check out the project on GitHub: roedyrustam/vibes-plug

πŸ“¦ NPM Package: vibes-plug on npm


What are your thoughts on multi-agent AI architecture and formal SMT verification for AI-generated code? Let's discuss in the comments below! πŸ‘‡

Top comments (0)