SINGLE-Clarity — One brain. Your work. $0 per cycle. 100% local.
11 agents. 1,885+ cycles logged to Notion. Fully autonomous. Watch it run in real time:
NEXUS ULTRA is a self-evolving, multi-agent AI swarm built on the SINGLE-Clarity architecture. It runs entirely on local hardware — no cloud, no API costs, no subscriptions.
11 specialized agents collaborate in timed cycles, scouting live signals from Reddit and HackerNews, writing outreach copy, critiquing each other's outputs, and logging every cycle to Notion via MCP in real time.
GENERATOR tier → COMMANDER · SCOUT · COPYWRITER · CONVERSION_ANALYST
CRITIC tier → VALIDATOR · SENTINEL_MAGNITUDE · METACOG · EXECUTIONER
OPTIMIZER tier → SUPERVISOR · REWARD · CLOSER
↑ |
└──────────── scores, lessons, memory, KG injection ──────┘
Every cycle, the REWARD agent scores performance. Top lessons are promoted into the next cycle's context. The swarm rewrites its own operating instructions based on what works.
SINGLE-Clarity is the cognitive system powering NEXUS ULTRA. It is not a framework or SaaS product — it is a unified local brain with five layered organs.
| One Brain | Single source of truth across all agents — nexus_kg.json |
| Your Work | Runs locally, $0 cloud, no API dependency |
| Self-Calibrating | 1,885+ cycles in production — smarter every run |
| Organ | Role | Description |
|---|---|---|
| KG | Memory | Knowledge Graph — typed, time-aware, 9,000+ nodes |
| CHRONOS | Brain | Temporal confidence engine — half-lives and decay |
| Swarm | Nervous System | 11 agents, 3 tiers, self-scored cycles |
| VeilPiercer | Immune System | Divergence detection, session tracing, FAILURE_MEMORY |
| NeuralMind | Interface | Force-directed KG graph, swarm health display |
| NEXUS ULTRA | ChatGPT / Claude | |
|---|---|---|
| Your prompts stay private | ✅ | ❌ sent to servers |
| Works with no internet | ✅ | ❌ |
| Monthly cost | $0 | $20+/mo |
| Learns from your sessions | ✅ persistent KG | ❌ resets |
| Multi-agent reasoning | ✅ 11 agents | ❌ single model |
| Notion live reporting | ✅ via MCP | ❌ |
1. Install Ollama and pull models
ollama pull qwen3:8b
ollama pull phi4-mini-reasoning
ollama pull llama3.1:8b2. Install Python dependencies
pip install httpx requests python-dotenv psutil3. Configure .env
cp .env.example .env
# Add your NOTION_TOKEN and NOTION_DATABASE_ID4. Launch the swarm
python nexus_swarm_loop.pyEvery swarm cycle is logged to Notion in real time via the Model Context Protocol.
What gets pushed (every ~35 seconds):
- 🔄 Cycle score, MVP agent, cycle type
- 🏆 Agent leaderboard — all 11 agents scored per cycle
- 🎯 Buyer intelligence signals from Reddit/HN scout
Setup:
- Create a Notion integration at notion.so/my-integrations
- Add your token and database IDs to
.env:
NOTION_TOKEN=ntn_your_token_here
NOTION_CYCLES_DB=your_database_id
NOTION_AGENTS_DB=your_agents_db_id
NOTION_BUYERS_DB=your_buyers_db_id
- Run the sync services:
python nexus_notion_sync.py
python nexus_notion_reporter.pyVeilPiercer exposes per-step agent tracing as native tools for Claude Desktop via the Model Context Protocol.
Edit %APPDATA%\Claude\claude_desktop_config.json:
{
"mcpServers": {
"veilpiercer": {
"command": "python",
"args": ["path/to/nexus-ultra/mcp/server.py"]
}
}
}Restart Claude Desktop. VeilPiercer appears in the tools panel.
| Tool | What it does |
|---|---|
start_session |
Start a new trace session for an agent run |
trace_step |
Log one agent step — captures prompt in, response out |
diff_sessions |
Compare two sessions — returns fork step and divergence |
pip install veilpiercer # free for local use→ PyPI · MCP Setup Guide
MIT — do whatever you want with it.
Built on: Ollama · Python · Qwen3 · Phi-4 · Llama · Notion MCP
SINGLE-Clarity architecture · March 2026