jonhill90/vibes

Vibes - Conversational Development Environment

Vibes transforms Claude Desktop into a conversational development environment through distributed MCP servers. Instead of learning command-line tools, you describe what you want to build and Claude implements it while teaching you.

Core Philosophy

Ask → Build → Understand → Improve → Create

Current Architecture

Production Stack: Claude Desktop + MCP Servers + Docker Containers + Vector Database

Vibes runs as a distributed system of specialized MCP servers, each handling specific capabilities:

Quick Start

Prerequisites

• Docker Desktop
• Claude Desktop
• Git

1. Clone Repository

git clone https://github.com/jonhill90/vibes.git
cd vibes
docker network create vibes-network

2. Start MCP Servers

cd mcp/mcp-vibes-server
docker-compose up -d
cd ..

3. Configure Claude Desktop MCP Settings

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "vibesbox": {
      "command": "docker",
      "args": ["exec", "-i", "mcp-vibesbox-server", "python3", "/workspace/server.py"]
    },
    "basic-memory": {
      "command": "docker",
      "args": ["exec", "-i", "basic-memory-mcp", "/app/start.sh"]
    },
    "MCP_DOCKER": {
      "command": "docker",
      "args": ["mcp", "gateway", "run"]
    },
    "archon": {
      "command": "npx",
      "args": ["mcp-remote", "http://localhost:8051/mcp"]
    }
  }
}

4. Restart Claude Desktop

After updating the configuration, restart Claude Desktop to load the MCP servers.

Current Capabilities

• Shell execution & container management via mcp-vibes-server MCP

  • Run bash commands in isolated environments
  • Docker container lifecycle management
  • Network access to vibes-network

• Desktop automation & visual feedback via mcp-vibesbox-server MCP

  • VNC desktop environment (XFCE4, 1920x1080)
  • Screenshot capture for Claude's vision
  • Mouse/keyboard control (click, drag, type)
  • ARM64 Chromium browser automation

• Task & knowledge management via archon MCP

  • Task tracking (find_tasks, manage_task)
  • RAG search across documentation (2-5 keyword queries)
  • Project management and organization

• Persistent memory via basic-memory MCP

  • Local Markdown-based knowledge storage
  • Conversation context across sessions
  • Semantic linking and knowledge graphs

• Container orchestration via MCP_DOCKER gateway

  • Unified MCP server management
  • Security isolation and secrets management
  • Enterprise observability

Context Optimization

We compressed the context Claude sees by 59-70% without losing functionality.

File Sizes Achieved:

• CLAUDE.md: 107 lines (from 143, 25% reduction)
• Patterns: 47-150 lines each (target ≤150)
• Commands: 202-320 lines each (target ≤350)

Context Per Command:

• /generate-prp: 427 lines (59% reduction from 1044 baseline)
• /execute-prp: 309 lines (70% reduction from 1044 baseline)

Impact: ~320,400 tokens saved annually (assuming 10 PRP workflows/month)

Why this matters: Claude has an "attention budget"—more tokens doesn't mean better results. By compressing context, we improved both speed and accuracy.

See for detailed methodology.

Context Engineering & PRPs

Vibes uses Context Engineering to transform AI coding from generic code generation to production-ready implementation. Instead of prompting in the dark, we engineer comprehensive context that enables first-pass success.

What are PRPs?

PRP (Product Requirements Prompt) = PRD + Curated Codebase Intelligence + Agent Runbook

A PRP is a context engineering artifact that treats Claude like a competent junior developer who needs a comprehensive briefing. It includes:

• PRD: Clear goals, business value, success criteria
• Curated Context: Documentation URLs, file references, existing patterns, known gotchas
• Agent Runbook: Implementation blueprint, pseudocode, task list, validation gates

Naming Convention: PRPs follow a standardized naming convention documented in . Key rules:

• Use prps/{feature_name}.md (no prp_ prefix)
• Initial PRPs use INITIAL_ prefix (auto-detected)
• Directory structure matches feature name

PRP Workflow

1. Create INITIAL.md      → Describe your feature
2. /generate-prp          → Research and create comprehensive PRP
3. /execute-prp           → Implement with validation loops
4. Validation             → Tests pass, code quality verified
5. Done                   → Production-ready code

Directory Structure

vibes/
├── .claude/              # Context-engineered components (59-70% reduction)
│   ├── commands/         # Slash commands for Claude Code
│   │   ├── generate-prp.md    # 320 lines (59% reduction)
│   │   ├── execute-prp.md     # 202 lines (70% reduction)
│   │   ├── list-prps.md
│   │   └── prep-parallel.md
│   ├── patterns/         # Reusable implementation patterns
│   │   ├── README.md          # Pattern library index
│   │   ├── archon-workflow.md
│   │   ├── parallel-subagents.md
│   │   ├── quality-gates.md
│   │   └── security-validation.md
│   ├── agents/           # Subagent specifications
│   └── templates/        # Report templates
├── prps/                 # Per-feature PRP artifacts
│   ├── {feature_name}/   # Feature-specific directory
│   │   ├── planning/     # Research outputs
│   │   ├── examples/     # Concrete examples
│   │   └── execution/    # Implementation artifacts
│   └── templates/        # PRP templates
│       ├── prp_base.md          # Comprehensive base
│       ├── feature_template.md  # Standard features
│       ├── tool_template.md     # API integrations
│       └── documentation_template.md
├── CLAUDE.md             # 107 lines (project rules only)
└── mcp/                  # MCP server implementations
    ├── mcp-vibesbox-server/
    └── mcp-vibes-server/

Core Principles

1. Context is King: Include ALL necessary documentation, examples, and gotchas
2. Validation Loops: Provide executable tests the AI can run and fix iteratively
3. Information Dense: Use keywords and patterns from your codebase
4. Progressive Success: Start simple, validate, then enhance
5. One-Pass Implementation: Comprehensive context enables getting it right the first time

Available Commands

• /generate-prp <file> - Research codebase and create comprehensive PRP
• /execute-prp <prp> - Execute PRP with iterative validation
• /list-prps [status] - List PRPs by status (active/completed/archived)

Example Templates

Choose the right template for your task:

• Feature Template: Standard feature development
• Tool Template: API integrations and external tools
• Documentation Template: READMEs, API docs, guides

Learn More

• - Comprehensive template with all sections
• - Reusable implementation patterns
• Context Engineering Intro - Original philosophy

Philosophy: Context engineering is 10x better than prompt engineering and 100x better than vibe coding. Give Claude the context it needs to succeed.

Pattern Library

The .claude/patterns/ directory contains reusable implementation patterns extracted from the PRP system. These patterns enable consistent, high-quality implementations across all features.

See for complete pattern documentation and usage guidelines.

Future Vision

Phase 1: Observable Agent Execution (In Development)

• Real-time screen sharing of AI work (terminal, browser, Neovim)
• Agent-specific environments with persistent state
• Pause/resume execution controls

Phase 2: Team Collaboration

• Discord-like interface for human-AI teams
• Multi-user knowledge spaces
• Agent coordination visualization

Phase 3: Advanced Intelligence

• Cross-session agent learning and skill accumulation
• Intelligent task routing between specialized agents
• Predictive workflow assistance

Conversational development environment in production. Observable AI execution in development.