ForgeDesigner alpha: an AI assistant that builds static sites locally on your own machine

ForgeDesigner has shipped its first alpha: a desktop application for macOS (Apple Silicon) and Linux (x86_64) in which an AI assistant builds websites from plain-language descriptions, with a live preview and output saved as ordinary files on your own disk. The developer announced the release in a dev.to post on September 24, positioning the tool as an alternative to both hand-coding HTML and renting space on a hosted website builder — no hosting account, no lock-in, and nothing stored in a vendor's cloud unless you put it there.

How it works

You describe what you want — the developer's example is a landing page with a short intro and a contact section — and the assistant generates the pages and writes the Markdown content while a layout preview renders next to the editor. Results land as plain files under a workspace folder (~/ForgeWorkspace/...), and the first output target, a ForgeCMS plugin, produces pages as SML plus Markdown that can be pushed to a Git repository and served as a static site.

The app is built on Forge 4D, an open platform where applications are written in two small languages: SML, a declarative markup for UI and content, and SMS, a scripting language for logic and event handling. The entire designer interface — docks, file tree, editor and the assistant tab — is itself an SML/SMS app running on a native runner written in C++ on top of Godot 4.6.

What the assistant can do

According to the dev.to announcement, the assistant has four tools — list_files, read_file, write_file and run_command — scoped to the workspace folder. Its actions appear in a transcript as it works, the file tree updates as files appear, and the editor stays in sync: every keystroke is saved, and changes the assistant makes show up in the editor immediately.

Offline operation is the default. Out of the box the assistant runs against a built-in mock that costs nothing and sends no data anywhere. Connecting a real model takes one configuration change: create an OpenRouter key, place it in a forge_ai_keys.sml file with 600 permissions inside the app's data folder, and set the provider to OpenRouter with a list of chat models — the first is primary, the rest act as fallbacks. If a provider reports being overloaded, the request retries automatically, with a manual retry button available after that.

Security model

An agent that can write files and execute commands needs guardrails, and the developer lays out several. Permissions are declared up front — workspace files, running programs after asking, and network access to OpenRouter — and confirmed once at first start. Commands are executed as argv arrays rather than shell strings; a short allowlist (git, ls, cat, ssh, scp) runs directly, while anything else opens a dialog where you decide to allow it once or permanently. API keys live in a runner-owned data file and are never handed to the app or the model.

The developer is also candid about limits: prompt injection remains a genuine risk, because trusted commands can reach anything the user account can — a key reason the release is labelled an alpha.

Rough edges and what's next

The macOS build is Apple Silicon only and merely ad-hoc signed, so users must clear the quarantine attribute once before launching; the Linux build covers x86_64 only, and there is no Windows version yet. ForgeCMS is the sole output plugin, free OpenRouter models are sometimes overloaded or withdrawn without notice, and the UI still shows traces of its origins as an SML editor.

Planned additions include deploying sites straight from the app via Git push over SSH, further output plugins targeting HTML, Android, macOS apps and even books exported to PDF or EPUB, and finer-grained always-allow rules for commands like git and ssh. The code is open source under GPLv3, with a commercial license available, hosted on Codeberg.

Why it matters

Most AI website builders today are cloud services: the prompt, the generation and often the resulting site live on someone else's servers. ForgeDesigner inverts that arrangement — the model does the authoring, but the artifact is a folder of plain files you own and can edit by hand at any moment, with the option of running entirely offline through the mock assistant. The transparent tool use and the honest accounting of prompt-injection risk also make it a useful early data point for how local-first AI coding agents could be built, even at alpha quality.