Your LLM is a brain in a jar. Give it a nervous system.
A biological architecture for artificial minds. Sensation, reflex, memory, and action—organized into coherent loops that turn chatbots into organisms.
🌐 Website: UNDER CONSTRUCTION! mnehmos.github.io/mnehmos.multi-agent.framework
📖 Blog Post: From Chatbot to Organism
| Server | Layer | Description |
|---|---|---|
| Somatic | 62 tools for computer automation: files, screen, keyboard, mouse, batch ops | |
| Autonomic | Memory persistence, agent handoff, file locks, bug tracking | |
| Autonomic | RAG pipelines, document Q&A, semantic search | |
| Reflex | Schema validation, contract enforcement, type scaffolding | |
| Central | Academic paper search for Deep Research mode | |
| Game Engine | D&D 5e combat, characters, encounters |
┌─────────────────────────────────────────────────────────────────────┐
│ CENTRAL (BRAIN) - Cognition │
│ Orchestrator • Architect • Planner • Ask • Deep Research │
│ ↑ Receives sensory input ↓ Sends motor commands │
├─────────────────────────────────────────────────────────────────────┤
│ SOMATIC (BODY) - Voluntary Action │
│ Red Phase • Green Phase • Blue Phase • Code • Debug │
│ ↑ Tool results ↓ Tool invocations │
├─────────────────────────────────────────────────────────────────────┤
│ AUTONOMIC (SUBCONSCIOUS) - Background │
│ Memory • Synch MCP • Index Foundry MCP │
│ State persistence, retrieval, agent coordination │
├─────────────────────────────────────────────────────────────────────┤
│ REFLEX (SPINAL CORD) - Immediate. │
│ (UNDER CONSTRUCTION) │
│ Trace MCP • Schema Validation • Contracts │
│ Rejects bad inputs before cognition │
└─────────────────────────────────────────────────────────────────────┘
# Create .roo directory in your project
mkdir -p .roo
# Copy templates
cp templates/custom_modes.yaml .roo/
cp templates/universal/AGENTS.md .Files to copy:
templates/custom_modes.yaml→.roo/custom_modes.yamltemplates/universal/AGENTS.md→AGENTS.md(project root)
Copy templates/custom-instructions-for-all-modes.md content to your IDE's global instructions field.
Install tools to extend capabilities. See templates/tools/ for integration docs:
| Tier | Tool | Layer | Integration Guide |
|---|---|---|---|
| 2 | OODA MCP | Somatic | ooda-mcp.md |
| 3 | Synch MCP | Autonomic | synch-mcp.md |
| 3 | Index Foundry | Autonomic | index-foundry-mcp.md |
| 4 | Trace MCP | Reflex | trace-mcp.md |
Progressive enhancement—start simple, add capabilities as needed:
| Tier | Components | What You Get |
|---|---|---|
| 0 | Baseline templates | Toolless coordination with 13 modes |
| 1 | + Native tools | File I/O, command execution |
| 2 | + OODA MCP | Full computer automation (62 tools) |
| 3 | + Synch + Index Foundry | Memory persistence + RAG |
| 4 | + Trace MCP | Schema validation layer |
Each tier preserves all contracts from previous tiers.
| Mode | Purpose |
|---|---|
| 🔄 Orchestrator | Task decomposition, delegation, integration |
| 🏛️ Architect | System design, ADRs, contracts |
| 📋 Planner | Task maps, backlogs, dependencies |
| ❓ Ask | Clarification, research synthesis |
| 🔎 Deep Research | Multi-source investigation |
| 🔬 Deep Scope | Issue scoping |
| Mode | Purpose |
|---|---|
| 🔴 Red Phase | Write failing tests |
| 🟢 Green Phase | Minimal implementation |
| 🔵 Blue Phase | Refactor with green tests |
| 💻 Code | Complex implementation |
| 🐛 Debug | Root cause analysis |
| Mode | Purpose |
|---|---|
| 📁 Memory | Documentation, knowledge management |
| 🗄️ Index Foundry | RAG pipeline specialist |
All work follows Observe → Orient → Decide → Act:
- Observe 👁️ — Read the task, scan files, check tests
- Orient 🧭 — Map what exists vs what's needed
- Decide 🎯 — Choose approach with lowest risk
- Act ⚡ — Execute one step, validate, repeat
🔴 RED → Write failing tests
🟢 GREEN → Minimal implementation
🔵 BLUE → Refactor with green tests
Every completed task returns structured data:
{
"type": "task-completed",
"task_id": "auth-impl-001",
"from": "green-phase",
"status": "success",
"files_changed": ["src/auth/jwt.ts"],
"tests_run": ["npm test -- auth"],
"summary": "Implemented JWT validation"
}Templates for every major AI-assisted IDE:
| IDE | Template Location |
|---|---|
| Roo Code | templates/custom_modes.yaml |
| Claude Code | templates/claude-code/CLAUDE.md |
| Cursor | templates/cursor/rules/_global.mdc |
| Copilot | templates/copilot/copilot-instructions.md |
| Path | Description |
|---|---|
templates/custom_modes.yaml |
Mode definitions (13 modes with layer mappings) |
templates/custom-instructions-for-all-modes.md |
Global contracts (paste into IDE) |
templates/universal/AGENTS.md |
Universal agent contract (OODA + TDD + Scope) |
| Tool Integration | |
templates/tools/ooda-mcp.md |
OODA MCP integration (Somatic layer) |
templates/tools/synch-mcp.md |
Synch MCP integration (Autonomic layer) |
templates/tools/index-foundry-mcp.md |
Index Foundry integration (Autonomic layer) |
templates/tools/trace-mcp.md |
Trace MCP integration (Reflex layer) |
LLMs are stateless text predictors. They have:
- ✅ High intelligence
- ❌ No agency
- ❌ No persistent memory
- ❌ No sensory-motor integration
A nervous system provides:
- Closed loops of action and sensation
- Reflexes that reject bad inputs immediately
- Autonomic processes that maintain state
- Central coordination of specialized capabilities
Chatbot: User → LLM → Response
↑
└── (no feedback)
Organism: Environment → Senses → Brain → Motor → Environment
↑ ↓
└──────── Feedback ────────────┘
MIT