AI Agent System

Sanctum runs a multi-agent AI system split across two platforms. Five specialized agents operate on the VM, and one agent runs on the Mac. Each agent has a defined role, access to shared skills, and participates in a coordinated intelligence network.

Agent Roster

VM Agents

The VM hosts five agents, each with a distinct specialization:

Agent	Role	Responsibilities
Yoda	Main	General intelligence, coordination, primary user interface
Windu	Security	Network monitoring, threat assessment, security briefings
Qui-Gon	Efficiency	Resource optimization, automation suggestions, cost analysis
Cilghal	Health	Genomic analysis, health monitoring, wellness insights
Mundi	Finance	Financial tracking, budget analysis, investment monitoring

All VM agents run through the OpenClaw gateway as a systemd user service. The gateway manages agent lifecycle, message routing, and plugin access.

Mac Agent

Agent	Role	Responsibilities
Jocasta	CRM	Contact management, relationship intelligence

Jocasta runs through the DenchClaw gateway on the Mac (port 18789) with its own configuration profile. It has direct access to Mac-native services and the Affinity CRM integration.

Model Configuration

Agents use a tiered model strategy with automatic fallback to maintain availability during outages or rate limits.

Primary Model

All agents default to Claude as the primary model via the Anthropic API. This provides the highest quality reasoning for complex tasks.

Fallback Chain

When the primary model is unavailable, agents fall through a chain of alternatives via OpenRouter and local LLMs:

1. Claude (Anthropic API) — primary
2. Qwen 3.5 Plus (OpenRouter) — first fallback
3. DeepSeek V3.2 (OpenRouter) — second fallback
4. MiniMax M2.5 (OpenRouter) — third fallback, with Presidio PII scrubbing
5. Local LLM (LM Studio, port 1234) — final fallback, fully offline

Note

OpenRouter fallbacks that route through non-US providers use Presidio PII scrubbing to strip personally identifiable information before the prompt leaves the local network.

Local Models

LM Studio runs on the Mac Mini (port 1234) serving local models for offline operation and low-latency tasks:

Model	Use Case
Qwen 3.5 35B (A3B)	General reasoning fallback

The Council-27B MLX server (port 8899) provides a Qwen 3.5 27B model with LoRA adapters fine-tuned for each of the six agents, enabling personalized responses at the local tier.

LiteLLM Proxy

A LiteLLM proxy (port 4001) provides an additional fallback path. When Anthropic returns billing errors, requests are automatically rerouted to OpenRouter through the proxy.

Gateway Architecture

Each platform runs its own gateway instance:

Platform	Gateway	Port	Bind	Config Location
Mac	DenchClaw	18789	LAN	~/.openclaw/
VM	OpenClaw	varies	LAN	~/.openclaw/

The VM gateway runs as a systemd user service with linger enabled, starting automatically at boot. It uses a SOPS wrapper script to decrypt secrets before launch.

The Mac gateway runs as a LaunchAgent. Both gateways use the same OpenClaw codebase but with independent configurations.

Council Bridge

The Council Bridge enables communication between Jocasta (Mac) and Yoda (VM). Since gateways block plaintext WebSocket connections to non-loopback IPs, all cross-instance communication routes through SSH.

Mac to VM

Jocasta → SSH to ubuntu@10.10.10.10 →
  openclaw agent --agent main --message "..."

VM to Mac

Yoda → SSH to bert@10.10.10.1 →
  PATH=<node path> openclaw agent --agent main --message "..."

Caution

VM-to-Mac SSH requires an explicit PATH prefix because the fnm-managed Node.js installation is not in the default SSH PATH. The bridge handles this automatically.

The bridge skill maintains a heartbeat every 2 hours to confirm bidirectional connectivity. If the heartbeat fails, an alert is raised through the watchdog system.

Plugins

Agents have access to shared plugins that extend their capabilities:

Plugin	Description
Supermemory	Persistent memory across conversations
Neo4j KG	Knowledge graph via Graphiti (port 18093)

Plugins are configured at the gateway level and available to all agents on that gateway instance.

Skills System

Agents access executable skills from the shared skills repository. Skills provide domain-specific tools that agents can invoke during conversations. See the Skills Development guide for details on the skill system architecture.

Skills are loaded from:

• The built-in skills directory in the OpenClaw installation
• Extra skills directories configured in openclaw.json (pointing to ~/Projects/openclaw-skills)

Holocron Chat Interface

The Holocron is a family-facing chat interface running on the Mac at port 19001. It uses a separate DenchClaw profile (~/.openclaw-dench/) with token authentication and LAN binding.

Setting	Value
Port	19001
Hostname	holocron (Firewalla DNS), holocron.local (mDNS)
Auth	Token-based
Access	http://holocron/ from LAN

The Holocron provides a simplified interface for household members to interact with the agent system without needing direct gateway access.