Personal Project Showcase

AI & Innovation

A passion project demonstrating the convergence of AI, infrastructure automation, and creative engineering.

15
Agents & Roles
490+
MCP Tools
45
Tool Groups
6
Platforms

What this proves

  • Production-style operations discipline: GitOps on a 9-node K3s cluster (~245 pods), observability by default (every service ships Prometheus metrics, alert rules and a Grafana dashboard), and deploys that wait for in-flight work to drain.
  • Security engineering applied to AI agents: scoped per-agent tool access, a 3-node OpenBao HA credential vault, a read-only tool loop for the local model, and an adversarial review stage before changes ship.
  • Platform engineering end to end: a tiered auto-remediation engine with an LLM risk judge and circuit breakers, an LLM failover chain (Claude → Codex → self-hosted Qwen on a DGX Spark), and Cloudflare WAF automation across 18 zones driven by Wazuh detections.

Meet Thor

A Self-Hosted, Multi-Agent AI Operations Platform

Thor is a comprehensive AI platform built from the ground up to manage, monitor, and automate an entire homelab infrastructure. Built on the Model Context Protocol (MCP), Thor orchestrates 13 specialized AI agents plus 2 review roles — each with its own personality, scoped tool access, and area of expertise — working together as a coordinated team that files and hands off work through durable tickets, from security monitoring to media management.

What started as a personal automation project is now a production multi-agent system with 490+ tools, automatic failover between cloud and self-hosted language models, tiered auto-remediation, a knowledge base over millions of indexed chunks, and an Obsidian vault as its single source of truth.

System at a Glance

  • Multi-Provider LLMs
  • Model Context Protocol (MCP)
  • TypeScript / Node.js
  • 13 Agents + 2 Review Roles
  • 490+ Tools / 45 Groups
  • Autonomous, Self-Healing Operations

The Agent Team

Each agent has a distinct personality, scoped tools, and a defined area of responsibility — plus two review roles that keep the team honest.

⚡

Thor

Overseer & Chief of Staff
Opus

The central brain — handles most tasks directly with the full suite of 490+ tools across 45 groups, and delegates longer work to other agents as durable tickets.

  • Full tool access across all domains
  • Ticket-based delegation and orchestration
  • Autonomous monitoring and alerting
  • Git workflow — branches, commits, PRs
🛡️

Heimdall

Security Monitoring & Threat Detection
Sonnet

The guardian — monitors threats via SIEM, IPS, and NVR systems, and blocks attackers at both the home firewall and the Cloudflare edge.

  • SIEM alert triage and investigation
  • Network intrusion detection
  • Edge WAF blocking for public sites
  • Automated threat response
🔧

Jarvis

Homelab Infrastructure
Sonnet

The engineer — manages virtual machines, containers, Kubernetes clusters, networking, and all infrastructure components.

  • Proxmox VM/LXC management
  • K3s cluster operations
  • Network configuration
  • Infrastructure monitoring
🏠

Ada

Home Automation
Sonnet

The homekeeper — controls smart home devices, automations, lighting, and environmental systems through Home Assistant integration.

  • Smart device control
  • Automation creation and management
  • Lighting and environmental control
  • Sensor monitoring and alerting
✨

Minerva

Business Assistant
Opus

The strategist — a dedicated assistant for business operations including website management, content strategy, and digital presence.

  • WordPress site management
  • Content strategy and creation
  • Business analytics
  • Digital marketing support
🎬

Apollo

Media Stack Management
Sonnet

The curator — manages the entire media pipeline from content discovery through availability on the media server.

  • Media library management
  • Automated content pipeline
  • Quality management
  • Storage optimization
🧭

Odyssey

Research & Planning
Sonnet

The explorer — a research-focused agent for long-term planning, relocation research, and multi-factor decision support.

  • Location research and comparison
  • Cost of living analysis
  • School and healthcare research
  • Timeline and milestone tracking
📜

Chronicle

Family Coordination
Opus

The organizer — helps with family care coordination: shared calendars, appointments, and keeping household paperwork in order.

  • Calendar and schedule tracking
  • Appointment reminders
  • Document organization
  • Family logistics
🏹

Artemis

Family Activities & Scheduling
Opus

The coach — helps plan family activities, schedules, and day-to-day logistics.

  • Activity planning
  • Schedule coordination
  • Reminders and checklists
  • Family logistics
🔨

Forge

Work Productivity
Opus

The craftsman — a work-productivity assistant for task tracking, development help, and searching technical documentation.

  • Task management
  • Development assistance
  • Code review support
  • Knowledge-base search (RAG)
🗡️

Offsec

Offensive Security & Red Team
Sonnet

The adversary — conducts penetration testing, OSINT gathering, and red team operations from a dedicated security environment.

  • Network scanning and enumeration
  • Vulnerability assessment
  • OSINT and reconnaissance
  • Exploit research
🎯

Fortuna

Career Research
Opus

The scout — researches the job market, scores opportunities against a structured rubric, and keeps career conversations organized.

  • Opportunity sourcing and deduplication
  • Structured fit scoring
  • Tailored application drafts
  • Follow-up tracking
🪽

Hermes

Laptop Helper
Sonnet

The messenger — a laptop-side helper agent that runs a small set of allowlisted jobs through a lightweight agent on the device.

  • Allowlisted laptop jobs
  • Heartbeat and availability tracking
  • Offline-aware scheduling
  • Fixed jobs, no caller arguments
🧐

Critic

Adversarial Review (role)
Sonnet

A review stage, not a chat persona — automatically challenges pull requests, risky plans, and finished tickets, and returns an approve / changes-needed / reject verdict.

  • Automatic PR and plan review
  • Structured verdicts
  • Comments posted on pull requests
  • Cannot review its own work
📚

Librarian

Knowledge Hygiene (role)
Sonnet

Keeps the knowledge vault healthy — fixes metadata, links notes from finished work, and flags duplicate notes and issues before anyone re-investigates a solved problem.

  • Nightly vault hygiene pass
  • Link and tag curation
  • Duplicate detection
  • Proposes, never deletes

How It Works

Thor is built on a modern, extensible architecture designed for reliability, observability, and safe autonomy.

Multi-Provider LLMs

Claude models for reasoning, with automatic failover to a second cloud provider and then to a self-hosted Qwen model on an NVIDIA DGX Spark — health-checked continuously, with circuit breakers

Model Context Protocol (MCP)

A single MCP server exposing 490+ tools in 45 groups, scoped per agent; the local model gets a curated, read-only tool loop

TypeScript / Node.js

Monorepo with a shared core library, MCP server, scheduler, Discord/Telegram bots, and a Hono + HTMX web portal

Durable Tickets

Delegated work becomes a ticket with a journal, heartbeat, staleness watchdog, and guaranteed delivery of the answer back to where it was asked

Obsidian Knowledge Vault

The vault is the source of truth for policy, runbooks, and agent memory; a watcher turns questions in notes into tickets and writes the answers back

RAG Knowledge Base

A RAG knowledge base indexes source code, technical documentation, and the Obsidian vault into Qdrant — roughly 2.9 million vector chunks searchable by every agent

Tiered Auto-Remediation

Alerts are classified, then auto-fixed, fixed after an objection window, or sent for approval — with verification, circuit breakers, and a remediation ledger

Credential Vault

A 3-node, highly available OpenBao cluster with auto-unseal, SSO login, and drift detection against the platform's own secrets

Git-Based, Drain-Safe Deploys

Agents ship through branches and pull requests; deploys wait for in-flight agent work to finish before restarting services

System Flow


  Discord / Telegram / Web / Obsidian / Mac / CLI
                     |
          [ Thor - Chief of Staff ]
                     |
       durable tickets + Critic review
       /     |      |      |      \
  Heimdall Jarvis  Ada  Apollo  Forge ...  (13 agents + 2 roles)
       \     |      |      |      /
      [ LLM chain: Claude -> Codex -> local Qwen ]
                     |
          [ MCP Tool Server ]
        490+ tools / 45 groups
                     |
  Proxmox | K3s | Home Assistant | Wazuh | Cloudflare
  UniFi | Synology | Grafana | GitHub | Obsidian | ...
        

Multi-Platform Access

Interact with the system from anywhere — chat, a web portal, your notes, or a laptop.

Discord

Per-agent channels, job threads you can reply to, and approval buttons

Telegram

Mobile-first access for on-the-go management

Web Portal

Tickets board, approvals inbox, and per-agent work history

Obsidian

Chat with any agent in a side pane, or just ask a question in a note

Mac & Laptop Agent

Native Mac app plus a lightweight laptop agent for on-device jobs

Claude Code CLI

Direct terminal interface for power users


Real-World Impact

Concrete capabilities of a system that runs every day.

Automated Security Operations

Detects and responds to threats autonomously — from SIEM alert triage to blocking attackers at the firewall and at the Cloudflare edge.

Infrastructure at Scale

Manages a 9-node Kubernetes cluster running ~245 pods via ArgoCD GitOps, with full lifecycle management.

Resilient by Design

Keeps working through provider outages by failing over between cloud and self-hosted models, with active health probes and breakers.

Self-Healing Infrastructure

Tiered auto-remediation fixes known issues on its own, asks before risky changes, and records every investigation in a ledger.

Agents That Review Agents

A dedicated Critic stage reviews pull requests and finished work, so autonomous changes get a second, adversarial pair of eyes.

Edge Observability

Cloudflare traffic, firewall events, tunnels, certificates, and domain renewals across every public zone feed Prometheus alerts and Grafana dashboards.

Hardened Secrets & Edge Defense

A 3-node OpenBao HA vault (Raft, auto-unseal) holds platform credentials, and Wazuh detections trigger blocks at the UniFi firewall, CrowdSec and the Cloudflare edge across 18 zones.

Monitoring & Observability

Every new service ships with Prometheus metrics, alert rules, and a Grafana dashboard — built and maintained by the agents themselves.


Technology Stack

The full technology landscape powering the Thor platform.

AI & ML

  • Claude AI (Anthropic)
  • Self-hosted Qwen (vLLM + LiteLLM)
  • Model Context Protocol
  • RAG on Qdrant

Backend

  • TypeScript / Node.js
  • Hono + HTMX
  • SQLite
  • Zod Validation

Infrastructure

  • Proxmox VE
  • K3s + ArgoCD GitOps
  • NVIDIA DGX Spark
  • OpenBao (HA)

Monitoring & Security

  • Wazuh SIEM
  • Grafana / Prometheus
  • Cloudflare WAF
  • CrowdSec / UniFi IPS

Automation

  • Home Assistant
  • Durable Tickets
  • Tiered Auto-Remediation
  • Git Workflows

Platforms

  • Discord / Telegram
  • Web Portal
  • Obsidian Plugin
  • Mac App / Laptop Agent

Project Evolution

From a simple automation script to a production-grade multi-agent AI system.

Thor has been in continuous development, growing organically as new challenges presented themselves. Each phase built on the previous one, with the agents themselves contributing to their own evolution.

2024

Project Genesis

Initial concept — exploring Claude AI for homelab automation. Built first MCP tools for Proxmox and Docker management.

Early 2025

Multi-Agent Architecture

Expanded to multiple specialized agents. Added Heimdall for security, Jarvis for infrastructure, and Ada for home automation.

Mid 2025

Platform Expansion

Launched Discord bot, Telegram bot, and web dashboard. Implemented persistent memory system and cross-agent communication.

Late 2025

Autonomous Operations

Agents began operating autonomously — scheduled monitoring, proactive alerting, and self-healing capabilities.

Early 2026

Consolidation & Scale

Consolidated the agent team around clear specializations and expanded tool coverage across the whole infrastructure.

Spring 2026

Maturity

Production-grade system managing all aspects of a complex homelab — from security to media to home automation.

May 2026

Knowledge & Self-Healing

The Obsidian vault became the source of truth for agent memory and policy, a RAG knowledge base was added across code and docs, and tiered auto-remediation began fixing alerts on its own.

Summer 2026

Local AI & Failover

Added a self-hosted model on an NVIDIA DGX Spark and a claude → codex → local failover chain with health probes, circuit breakers, and a read-only tool loop for the local model.

September 2026

Chief of Staff

Durable tickets replaced fire-and-forget delegation; added the Critic and Librarian roles, a web portal for tickets and approvals, drain-safe deploys, and a 3-node OpenBao credential vault.

Fall 2026

Edge & Everywhere

Cloudflare WAF control and edge observability across every public zone, a laptop agent for on-device work, and a vault watcher that answers questions written in notes.

Want to know more?

Dive deeper into the architecture, agents, and real-world accomplishments.

Get in Touch