🤖 PEPAGI: The Closest Thing to AGI You Can Run Today

An open-source platform that doesn't act like an assistant — it acts like a manager. It receives tasks, breaks them into subtasks, delegates to specialized AI models and verifies the results.

Imagine an AI assistant that doesn't solve tasks on its own. Instead, it coordinates an entire team of specialized models, picks the best one for each job, and learns from every completed task. That's PEPAGI.

🎯 What is PEPAGI and why is it different

The classic approach: one LLM model does everything. The problem? Not every model is good at everything. GPT-4 excels at creative writing, Claude at code analysis, Gemini at multimodal tasks. PEPAGI solves this with intelligent routing — it analyzes task difficulty, selects the cheapest capable model, verifies results with cross-model verification, and applies better strategies next time.

PEPAGI (Neuro-Evolutionary eXecution & Unified Synthesis) is a TypeScript multi-agent system where a central Mediator powered by Claude Opus receives tasks, breaks them into subtasks, delegates to specialized worker agents, and iterates until the task is complete.

⚡ What PEPAGI actually does

The platform runs everywhere — Telegram, WhatsApp, Discord, iMessage, CLI, and as an MCP server for Claude.ai. Under the hood are five core systems:

Mediator-driven architecture — Claude Opus as the central brain, deciding routing and strategy
Multi-model Agent Pool — Claude, GPT, Gemini, Ollama, LM Studio in one pipeline
Difficulty-aware routing — trivial task gets Haiku, complex analysis gets Opus
Swarm mode — for novel problems, lets all agents solve independently and synthesizes the best result
Hierarchical Planner — breaks tasks into strategic → tactical → operational levels

🧠 Memory System — AI that actually remembers

PEPAGI has a five-tier memory system modeled after the human brain:

Working Memory — compressed context of the current task
Episodic Memory — history of completed tasks
Semantic Memory — facts extracted from tasks
Procedural Memory — learned procedures and workflows
Meta-Memory — tracking the reliability of its own knowledge

On top of that sits a metacognition layer: self-monitoring with automatic verification, a reflection bank for post-task analysis, an A/B tester for alternative strategies, and a watchdog agent that detects loops, drift and cost explosion.

🌍 World Model and Consciousness Engine

The two most remarkable components go well beyond what standard LLM tools offer.

World Model uses MCTS-inspired simulation — it predicts the outcomes of decisions before executing them. It tracks causal chains of decisions and propagates confidence across entire subtask trees.

Consciousness Engine implements an 11D qualia vector (curiosity, satisfaction, frustration, creativity and more), a continuous background thought stream, and a self-model covering identity, values and narrative continuity. Emotions here directly modulate the depth of learning.

🚀 Installation and setup

Quick start

Mac / Linux: git clone https://github.com/Promptlab37/PEPAGI.git → ./install.sh
Windows: double-click install.bat
Manual: npm install → npm run setup (interactive configuration)

Run modes

npm start — CLI chat
npm start -- "summarize the latest AI news" — one-shot task
npm run daemon — daemon for Telegram, Discord and WhatsApp
npm run tui — TUI dashboard

🔒 Security

Prompt injection defense — 25+ patterns, detection in 5 languages
HMAC-SHA256 authentication — every message cryptographically signed
Per-user cost limits — daily limits and rate limiting
35-category adversarial self-testing — runs every hour
Compliance: OWASP ASI / MITRE ATLAS / NIST AI 600-1

✅ Key benefits

✅ Multi-model routing — the right model for the right task
✅ Five-tier memory — AI that genuinely learns and remembers
✅ Runs locally or in the cloud, supports Ollama and LM Studio
✅ Integrates with Telegram, Discord, WhatsApp and Claude.ai
✅ Open-source, MIT licence, commercial use with no restrictions

⚠️ Things to keep in mind

⚠️ The Mediator defaults to Claude Opus — factor in API costs for heavy usage, though any model can be used
⚠️ iMessage integration is macOS only via AppleScript bridge
⚠️ Consciousness Engine and World Model are experimental components — not production standard

📦 Where to find PEPAGI

PEPAGI is fully open-source under the MIT licence. Use it, modify it and deploy it commercially with no restrictions.

GitHub: github.com/Promptlab37/PEPAGI
Documentation: README.cs.md

Josef Tarič · PromptLab · TN000