Video: "I Rebuilt Hermes in Claude Code (It's Ridiculously Good)" by Julian Goldie on YouTube.
What "rebuilding" actually means here
This isn't rewriting Hermes's source code. It's using Claude Code to build a new agentic system — a locally hosted dashboard with chat, memory, scheduling, and tool-calling — that replicates the core behaviour of Hermes from the outside. The distinction matters: you're directing Claude Code to design, build, and test a functional system by describing what you want, not by writing the code yourself.
The four-layer architecture Julian used puts Claude Desktop as the planner, Claude Code as the builder, an Obsidian vault as shared memory, and the resulting dashboard as the control surface. The whole thing was built in approximately one session, without Julian writing a line of traditional code.
What the Claude Code build process looked like
Claude Code takes terminal-level instructions and executes them. You describe the component you want — "add a Kanban board that syncs with my agent's task queue" — and Claude Code writes the logic, tests it, and integrates it. The session Julian showed covered a dashboard frontend, an API layer connecting to the Hermes memory store, and several utility skills the agent can call on demand.
In practice the process involves more back-and-forth than a single prompt. Bugs come up, edge cases need handling, and some components need re-specifying when the first attempt doesn't quite match the intent. But the overall arc — from blank project to working agent interface in hours rather than weeks — is genuinely new.
What this shows about AI-native development
The point of the experiment isn't "you can rebuild Hermes yourself". It's that the gap between a working prototype and a tool you've bought has nearly closed. If an off-the-shelf agent doesn't quite do what your business needs, you now have a practical path to a custom version — one that fits your specific workflow, stores context the way you actually work, and doesn't require a development agency to build.
That has wider implications. Businesses that understand how to direct Claude Code as a builder — rather than just as a writing assistant — are sitting on a capability most of their competitors don't have yet.
Where it still needs a human in the loop
Claude Code is a capable executor, but it benefits from a clear brief and someone who knows what "done" looks like. Vague instructions produce plausible-looking code that misses the point. And once you have something working, the ongoing maintenance — updating skills, adjusting memory schema, connecting new tools — needs someone who understands the structure well enough to direct changes sensibly.
For businesses without that background, the right move is to work with someone who already knows how these systems fit together, rather than learning from scratch on a live setup.
Where this connects to NordSys
We build custom Claude Code and Hermes Agent setups for businesses that need something more specific than the default install. Whether that's a particular memory structure, a bespoke skill library, or an integration with tools you already use — the approach Julian demonstrated here is exactly how we work. If you want something built to your workflow rather than adapted to ours, get in touch.
See our AI Agents service →