The Stack.No Smoke.
This isn't a list of tools I read about. This is what's running right now — the models, services, and infrastructure I use to build everything you see on this site.
Models
24 Ollama Cloud models, routed by task type
GLM-5.2
Ollama Cloud
Flagship — main conversation + heavy reasoning
Kimi K2.7 Code
Ollama Cloud
Coding sub-agents + delegate tasks
MiniMax M3
Ollama Cloud
Fast tier — quick lookups, simple tasks
Gemma 4
Ollama Cloud
Vision + image analysis
DeepSeek V4 Pro
Ollama Cloud
Fallback — coding
Nemotron 3 Super
Ollama Cloud
Fallback — reasoning
Qwen 3.5
Ollama Cloud
Fallback — multimodal
Services
All running on a single WSL2 box, 127.0.0.1 only
Hermes Gateway
:8642AI orchestrator + Telegram bot
Python, systemd
FCC Proxy
:8082Claude Code → Ollama Cloud routing
Python, uvicorn
Ollama
:11434Local + cloud model inference
Go, systemd
MemPalace
Persistent memory + knowledge graph
Python, MCP
GitNexus
Code knowledge graph (60K+ nodes indexed)
Node.js, CLI
Open WebUI
:3000Browser-based chat interface
Docker
Toolchain
What each layer does
Orchestration
- ▸Hermes Agent
- ▸delegate_task subagents
- ▸Kanban multi-agent boards
Coding
- ▸Claude Code (via FCC proxy)
- ▸GitNexus code graph
- ▸Superpowers skills
Memory
- ▸MemPalace MCP
- ▸Semantic search
- ▸Knowledge graph
- ▸Session diary
Models
- ▸21 Ollama Cloud models
- ▸NVIDIA NIM (fallback)
- ▸7-model fallback chain
Infrastructure
- ▸WSL2 Ubuntu
- ▸Systemd services
- ▸Docker containers
- ▸iptables firewall
How It Actually Works
A message comes in via Telegram. Hermes (the orchestrator) reads it, decides what to do, and either handles it directly or delegates to a coding subagent running through the FCC proxy. MemPalace remembers everything across sessions. GitNexus maps the codebase so changes are surgical, not guesswork. All inference runs through Ollama Cloud — 24 models, no per-token costs.
Config backed up to github.com/joblas/Lurkr-Jo-Blade-Hermes · Last sync: July 7, 2026