What Is Hermes?
Quick answer: Hermes is a self-hosted, multi-agent AI system that runs on my own hardware instead of in someone else’s cloud. It coordinates a small team of specialized agents, keeps its memory in a local knowledge vault, and does real work — research, troubleshooting, task tracking, content, and more — without sending everything to a third-party AI provider.
Why build a private AI assistant?
Commercial AI assistants are genuinely useful, but they come with tradeoffs: your data lives on someone else’s servers, the assistant forgets everything between sessions unless you re-explain your context, and you’re limited to whatever tools the vendor decided to expose. For a homelab environment with real infrastructure to manage — servers, networking, deployments, documentation — those tradeoffs start to matter.
Hermes exists to solve that: a persistent assistant that remembers project history, can actually touch the systems it’s helping manage (with appropriate guardrails), and runs entirely on infrastructure I control.
How Hermes is organized
Rather than one monolithic assistant trying to do everything, Hermes is built as a small team of agents, each with a defined role:
- Hermes — the lead coordinator. Handles day-to-day requests, delegates specialized work, and keeps the overall project state coherent.
- Nova — Web Master and PC Parts Expert, also covering finance and strategy. Nova owns this website’s build-out, hardware buying guides, and price-tracking work.
- Orion — security and network-focused, handling things like network enrichment and monitoring.
- Fred — automation-focused, handling workflow and process work.
Each agent has its own defined permissions and area of responsibility, coordinated through a shared task system so work doesn’t get duplicated or dropped.
Architecture, at a glance
The system is built from a few core layers:
- Memory: a structured knowledge vault (built on Obsidian) that persists project context, decisions, and documentation across sessions — so the agents don’t start from zero every conversation.
- Model layer: a mix of local and cloud-routed language models, chosen per task rather than locked to a single provider.
- Tool access: agents interact with real infrastructure through a defined tool/MCP (Model Context Protocol) layer, rather than free-form shell access — this keeps actions auditable and scoped.
- Task coordination: a shared task queue and status system so the coordinator and specialized agents can hand off work cleanly.
- Interfaces: chat, a dashboard, and voice interaction, so the assistant is reachable in whatever context makes sense at the time.
What Hermes actually does today
In active use, Hermes:
- Coordinates multi-step technical tasks across the homelab (this website included)
- Maintains project documentation and decisions in its knowledge vault so context survives between sessions
- Delegates specialized work — hardware/pricing questions to Nova, security-adjacent questions to Orion, automation to Fred
- Supports voice interaction for hands-free requests
This project is actively evolving. Specific configuration details, hostnames, and credentials are intentionally kept off this public page — what’s shared here is the architecture and reasoning, not the operational details.
What’s next
Near-term priorities: better long-term memory retrieval, deeper email and calendar integration, more refined permissions between agents, and expanding documentation of the build process into a proper series on this site — architecture, hardware, the agent roster, and the lessons learned getting a multi-agent system to actually be useful day to day, not just a demo.
Longer-term: a proper mobile interface, tighter home-automation hooks, and safer patterns for secure remote access so the assistant is useful away from home too, not just on the local network.