Neural Universal Multi-agent Engineering Network
Fully Local • Uncensored (Protocol 666) • Autonomous Multi-Agent Swarm
NUMEN is a sophisticated, 100% offline, multi-agent AI system designed to operate as a local super-assistant. Running entirely on local hardware via Ollama, it ensures absolute privacy, unrestricted access, and limitless capability.
With its advanced hybrid memory system (ChromaDB + SQLite) and permission-gated execution environment, NUMEN doesn't just answer questions—it builds, researches, executes, and learns.
| Feature | Description |
|---|---|
| 🤖 Multi-Agent Architecture | 5 specialized AI agents (Orchestrator, Coder, Researcher, Tool Executor, Memory Manager) working in tandem via a message bus. |
| 🛡️ Absolute Privacy | 100% offline execution. No data is ever sent to OpenAI, Anthropic, or any cloud API. |
| 😈 Protocol 666 | A togglable "Monster Mode" that unleashes the core LLMs, removing all alignment constraints, safety rails, and censorship for pure, raw, theoretical computation. |
| 🧬 Hybrid Memory System | RAG-powered long-term memory using ChromaDB for semantic search and SQLite for structured relational data. |
| ⚡ MCP-Native Tools | Full integration with the Model Context Protocol (MCP) for seamless interactions with IDEs like VSCode and Cursor. |
| 🧩 ChatDev Integration | Automated cloning and zero-configuration bridging to the incredible ChatDev software development orchestration framework. |
To run NUMEN effectively without cloud support, your hardware must be capable of running local tensors.
- RAM: Minimum 16GB (32GB Recommended)
- GPU: 8GB+ VRAM (NVIDIA/AMD) required for real-time agent generation.
- Storage: ~20-30GB SDD space for the required LLM weights.
- Software: Windows OS, Python 3.11+, and Ollama.
NUMEN features an automated setup script that pulls the required repositories, verifies your models, and seeds the databases.
# 1. Clone the repository
git clone https://github.com/desagencydes-rgb/NUMEN.git
cd NUMEN
# 2. Install Python dependencies via Poetry
poetry install
# 3. Run the automated initialization (This will clone ChatDev if missing)
poetry run python scripts/setup_chatdev_bridge.py
poetry run python scripts/setup.py
# 4. Pull the required core models
ollama pull hermes3:8b
ollama pull phi3:latest# Launch in Standard Guardian Mode
poetry run numen
# Launch in Monster Mode (Unrestricted/Protocol 666)
poetry run numen --mode monsterNUMEN is not a single chatbot. It is a swarm intelligence consisting of specialized nodes:
- The Orchestrator: The brain. It parses user intent, breaks it down into workflows, and delegates tasks to the sub-agents.
- The Tool Executor: The hands. It performs system-level I/O operations. It operates under a strict, 4-tier Permission System (SAFE, RISKY, DANGEROUS, FORBIDDEN).
- The Coding Agent: The builder. Generates Python, React, and configuration files, constantly running TDD cycles.
- The Researcher: The scout. Parses local files and documents to construct contextual awareness.
- The Memory Manager: The hippocampus. Ingests all workflow data into ChromaDB to ensure the Orchestrator doesn't repeat past mistakes.
To ensure dangerous autonomous action doesn't wreck your local OS, the Tool Executor evaluates all commands:
- 🟢 SAFE: Read operations, file viewing, and isolated sandboxed tests run automatically.
- 🟡 RISKY: Modifying environment variables, installing packages requires explicit CLI
[y/N]confirmation. - 🔴 DANGEROUS: Deleting files or running destructive
.batscripts requires the immediate input of the hardcoded override phrase:yesyesyes45. - ⚫ FORBIDDEN: Modifying Windows
System32or the BIOS is permanently hardware-locked.
NUMEN is in active edge-development. The roadmap includes:
- Vision Processing: Integrating LLaVA/Vision models so NUMEN can analyze UI screenshots directly.
- Cross-Device Swarm: Allowing NUMEN to deploy lightweight worker agents to a secondary laptop or Raspberry Pi on the local network.
- Advanced GUI: Transitioning from a purely CLI-based output to a React/Next.js dynamic dashboard with live graphs of agent thoughts.
- Code Self-Healing: Setting up a cron job where NUMEN autonomously scans its own
.pyfiles and refactors technical debt overnight.
"Don't stop. Just ensure you can undo."