Building AETHUR: a Claude Code-style design platform on Hermes

Building AETHUR: a Claude Code-style design platform on Hermes

Developers got Claude Code — an agent that lives in the repository, reads the codebase, runs the tests and ships reviewed work. Design, meanwhile, mostly got a chat box: describe a hero section, receive a PNG, start again tomorrow with an assistant that remembers nothing.

We think that gap is the actual opportunity, so we are building AETHUR: an AI design platform interface that runs on Hermes Agent and behaves like a colleague rather than a vending machine. This is an honest look at what exists today, what the architecture looks like, and what it teaches anyone building things for AI agents to use.

What AETHUR actually is right now

AETHUR lives in its own project folder with its own Git history, and we will not pretend it is further along than it is. The current commit message is "Build AETHER workspace interface" — singular, because the core is one serious workspace screen.

The stack under it:

• React 19 rendered through vinext on Vite 8, giving us server components with fast iteration
• Tailwind CSS 4 and TypeScript throughout, linted with accessibility and hook rules switched on
• A conversation API route that carries dialogue between the human and the agent layer
• A Cloudflare Worker deploy target, with Drizzle/D1 available when persistence is genuinely needed
• A rendered-HTML test that checks what actually ships, not just what compiles

That last point matters more than it sounds. If an AI is going to touch an interface, you need deterministic proof of what changed. Tests on rendered output are how a supervised agent earns trust commit by commit.

Why running it on Hermes changes the maths

A stateless chat window forgets your brand rules every session. Hermes gives an agent two things a chat box cannot: skills and persistent memory.

Per the Hermes documentation, skills are on-demand knowledge documents the agent loads when a task matches — procedural memory for how we do things here, from design conventions to publishing gates. Memory persists across sessions within explicit character budgets, so the agent accumulates judgement about a project instead of starting cold each morning.

For a design platform that means AETHUR can hold house style, component decisions and review standards as durable context — and improve its own procedures when work reveals a better path.

Supervision is the product

The premium version of AI-assisted design is not a bigger model. It is governance. AETHUR is built around supervised operation: scoped permissions, approval gates before anything consequential happens, and a human reviewer signing off material decisions.

That is also our answer to the race to the bottom. Cheap AI output is abundant. Judgement, taste, speed and accountability are not — and a supervised design agent is a tool for the second list, not the first.

What it teaches about building agent-ready anything

There is a direct line from this build to our AEO work. If you want AI agents to operate on your website or product, give them a structured, inspectable environment:

1. Clear contracts — defined routes and APIs, like AETHUR's conversation endpoint, beat implied behaviour
2. Testable surfaces — rendered-output checks make agent edits verifiable, the same way schema makes content machine-checkable
3. Documented conventions — skills files beat tribal knowledge, for agents and new hires alike

The discipline that makes a site legible to answer engines is the same discipline that makes a codebase operable by an agent. Structure is structure.

What this does not prove yet

AETHUR is a work in progress. We are not claiming production metrics, client outcomes or revenue effects from it — the honest status is: real architecture, real tests, early days. Follow along as it matures.

The bet is simple: the next generation of design tools will be supervised agents with memory, operating inside well-structured environments. We would rather build that now than demo a chat box.