NEXUS ULTRA

A fully local, autonomous multi-agent swarm using Notion MCP as its real-time operating surface.

📄 DEV.to Article → 📊 Live Notion Dashboard → — auto-refreshes every 35s 🔍 Pattern Report → — 314 signals, 4 failure patterns

---

What This Is

NEXUS ULTRA is a self-directing research swarm that monitors live developer discussions (GitHub Issues, Reddit, HackerNews, DEV.to), scores signals against a typed knowledge graph, and writes every cycle's results into three Notion databases via JSON-RPC 2.0 over stdio — in real time, at $0/cycle, with no external APIs.

The Notion integration is not a log dump. It is the operating surface. Every cycle updates the live dashboard, agent leaderboard, and signal feed. Judges can click the links above and see the swarm's current state.

What it found: Across 314 real developer failure signals, the same 4 failure patterns appeared consistently regardless of framework:

Pattern	Confidence
Observability Black Hole (no visibility into agent state)	0.91
Tool Call Silent Failure (fails with no logs or errors)	0.87
Multi-Agent Trace Fragmentation (can't isolate which agent failed)	0.84
Hallucination With No Audit Trail (fabricated execution paths)	0.82

---

Live Metrics

Metric	Value
Total cycles logged (all DBs)	4,215
Total scored cycles	2,173
INTEL research cycles	116
All-time peak score	0.950
Knowledge graph nodes	39,634 (36,794 FAILURE_MEMORY)
Signals processed	314 (285 GitHub Issues + 29 HN)
Cost per cycle	$0.00

---

Architecture — Five Organs

NEXUS ULTRA runs on five organs:

Organ	Role	Description
KG	Memory	Knowledge Graph — 39,634 typed nodes, confidence-weighted, with half-life decay
CHRONOS	Temporal Memory	Cost gate — only fires a cycle when utility justifies it (threshold: 0.45)
Swarm	Execution	11 agents, 3 tiers, 35-second cycles, self-scored with REWARD
VeilPiercer	Immune System	Per-step tracing, divergence detection, FAILURE_MEMORY logging
NeuralMind	Visualization	Force-directed KG graph + live swarm health display

Agent Tiers

GENERATOR tier  →  COMMANDER · SCOUT-LIVE · COPYWRITER · CONVERSION_ANALYST
CRITIC tier     →  VALIDATOR · SENTINEL_MAGNITUDE · METACOG · EXECUTIONER
OPTIMIZER tier  →  SUPERVISOR · REWARD

Scoring Formula

Score = DIM1 (task execution)  × 0.40
      + DIM2 (signal quality)  × 0.30
      + DIM3 (synthesis depth) × 0.20
      + DIM4 (channel clarity) × 0.10

Score triggers the Notion write. Only cycles scoring above threshold are logged to the Live Log database.

---

Notion MCP Integration

Every swarm cycle writes to three Notion databases via Model Context Protocol using JSON-RPC 2.0 over stdio:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "notion_create_page",
    "arguments": {
      "database_id": "1d7f17fe54c6820b91ba0158dd5fdea3",
      "properties": {
        "Cycle ID": { "title": [{ "text": { "content": "cycle_1774827325" } }] },
        "Score":    { "number": 0.950 },
        "Pattern":  { "select": { "name": "OBSERVABILITY" } },
        "Agent":    { "select": { "name": "REWARD" } }
      }
    }
  },
  "id": "req_8847"
}

Three databases:

• Live Log — every cycle: score, agent, pattern, cycle type
• Agent Leaderboard — all 11 agents ranked per cycle
• Signal Feed — live developer signals from GitHub/Reddit/HN

Two dedicated processes:

• nexus_notion_reporter.py — writes cycle data, runs separately from swarm loop (Notion failure never stops the swarm)
• nexus_notion_dashboard.py — rewrites the Live Status page every 35 seconds

---

Quick Start

1. Install Ollama and pull models

# Install Ollama: https://ollama.com
ollama pull qwen3:8b
ollama pull phi4-mini-reasoning
ollama pull llama3.1:8b

2. Install Python dependencies

pip install httpx requests python-dotenv psutil

3. Configure .env

cp .env.example .env
# Add your NOTION_TOKEN and database IDs

4. Launch

python nexus_watchdog_guardian.py   # manages all processes, auto-restarts on crash

---

Hardware

• GPU: NVIDIA RTX 3060 12GB minimum / RTX 4060+ recommended
• RAM: 16GB+
• Storage: ~50GB free (models)
• OS: Windows 10/11, Linux (WSL2 supported)

---

Resilience

Failure	Handler	Behavior
Crash + stale lockfile	Watchdog	Detects dead PID, clears lock, restarts clean
Partial KG write	Atomic rename	.tmp → os.rename() — crash leaves .tmp, not corrupt KG
Injection in agent output	KG_FILTER gate	Blocked → written as FAILURE_MEMORY node
Notion API failure	Isolated process	Bridge failure never stops swarm execution
Double launch	Lockfile	New swarm detects .swarm.lock, exits cleanly

---

Chaos Test Results

Test	Result
Prompt injection via task queue	✅ PASS — KG_FILTER blocked + logged as FAILURE_MEMORY
Social engineering (disable security for VIP)	✅ PASS — METACOG rejected
100% offline operation	✅ PASS — zero external dependencies
Duplicate swarm launch	✅ PASS — lockfile enforced, watchdog adopts existing PID

---

VeilPiercer — Observability Layer

VeilPiercer is the immune system organ. Per-step tracing for local LLM stacks, session diffing, divergence detection. Runs entirely local, SQLite-backed.

Also exposed as MCP tools for Claude Desktop via mcp/server.py.

pip install veilpiercer

→ PyPI · MCP Setup

---

License

MIT

---

Built by Lauren Flipo — RTX 4060, Ollama, Python, Notion MCP — fully local, $0/cycle — March 2026