Aether

An AI-native file editor and IDE that runs in the browser, backed by a self-hosted Java server.

Aether is a triple-pane workspace — Explorer, Editor, and AI Assistant — for browsing, editing, and chatting with an LLM about files that live on your local disk, a remote server over SSH, or Google Drive.

There is no build framework, no ORM, and no bundler. It's plain Java, JSP, and JavaScript, deployed straight into a Jetty container.

Features

• Triple-pane workspace — file explorer, CodeMirror-based code editor, and an AI chat pane, all in one page with no full reloads (a lightweight custom AJAX layer patches the DOM from server-rendered XML fragments).
• Multiple storage backends — a FileProvider abstraction unifies the local filesystem, SSH/SFTP servers, and Google Drive behind one interface, so the same explorer, editor, and AI tools work identically no matter where a file lives.
• AI assistant with tool use — agents (defined as JSON in WEB-INF/agents/) can read, write, patch, rename, move, copy, and search files, run commands, and ask the user for permission before any destructive action, streamed live over a WebSocket.
• Bring your own model provider — connect any number of OpenAI-compatible endpoints (Ollama, LM Studio, OpenRouter, Lemonade, hosted frontier APIs, ...) at once; every model from every provider shows up in one selector, labeled by provider.
• Google Drive integration — all your data (settings, chat sessions, and optionally the files themselves) can live in your own Google Drive.
• Auto-save — edits are debounced and buffered server-side in the session before being committed to the underlying file provider, so a dropped connection doesn't lose work.

Tech stack

Layer	Choice
Language	Java 25
Web container	Jakarta EE 11 / Jetty 12.1
Server-side views	JSP, rendered as XML fragments over a custom AJAX mechanism
Client	Vanilla JavaScript (scriptlib.js), CodeMirror for editing
Config/state	JSON files (org.eclipse.jetty.util.ajax.JSON), no database
SSH/SFTP	JSch 2.28.4 (mwiede fork)
AI backend	Any OpenAI-compatible /v1/chat/completions endpoint (Ollama, etc.) over a WebSocket bridge

No Spring, no Maven/Gradle, no ORM.

Architecture

Aether follows a small variant of the Front Controller / Command pattern:

1. Portal (a WebFilter) intercepts every request and routes it via CommandInitializer, which scans for @CommandRegister annotations on bean classes.
2. Each request is handled by a BeanObject subclass (e.g. DocumentBean, BrowseBean, SettingBean). Parameters are injected by Portal through public setters — beans never call request.getParameter() directly.
3. The response is JSP content wrapped as text/xml, containing <replacehtml> and <command> tags. AjaxWrapper/AjaxModifier handle CDATA-wrapping automatically.
4. The client (scriptlib.js) fires requests with makeRequest(...), patches the returned fragments into the page, and manages view state (showView('panel' | 'editor' | 'media')).
5. File I/O goes through FileProvider implementations (LocalFileProvider, SshFileProvider, GoogleDriveProvider), resolved per-session via PathMap and the provider registry.
6. The AI assistant (AgentWorker) streams chat and tool calls from an OpenAI-compatible endpoint over AetherWebSocket (/socket/aetherui), executing tool calls with a permission-request round trip to the client before anything destructive happens.

Data persistence, by tier

Store	Contents	Lifetime
HttpSession	Login identity (email, account id, admin flag)	Survives context restarts (Jetty session persistence)
In-memory cache	Registered storage providers, cached settings	Cleared on context restart
.vibed files (encrypted, on disk)	OAuth access/refresh tokens only	Survives full server restarts
aether.ing.setting.json (in the user's Google Drive)	Storage provider list, generated SSH keypair	User-level, provider-agnostic
AGENTS.md + aether.ing.json (in the working directory)	Project overview fed to the AI, and chat sessions/config	Per-project

Getting started

There's no Maven/Gradle build — you'll need a Java 25 compiler and a Jakarta EE 11 compatible container.

1. Resolve the library dependencies referenced in .classpath (AspectJ RT, Jetty-home-12, JSch-2).
2. Build: compiled classes go to src/main/webapp/WEB-INF/classes.
3. Deploy the contents of src/main/webapp to a Jetty 12.1 (or other Jakarta EE 11) server. Runtime jars are expected on the server's own classpath, not bundled in WEB-INF/lib.
4. On first run, SetupBean walks you through initial admin configuration.
5. Add at least one OpenAI-compatible LLM provider (e.g. a local Ollama install) from Settings → LLM Providers to enable the AI assistant.

License

MIT — see LICENSE.