Describe your site in plain language - purpose, audience, style, content. An autonomous agent researches the topic, designs and writes the complete page as a single self-contained index.html on Bootstrap 5. It runs entirely on your own hardware against your own LLM.
Autonome is a handful of files you can keep in any folder. No npm, no build step, no external tracking. Here's what that small footprint does.
The agent doesn't just complete text. It researches your topic - optionally via Wikipedia search tools - reasons about the brief, and writes a complete, responsive website in one pass.
search_wikipedia · get_wikipedia_page → one html blockWorks with LM Studio or Ollama (OpenAI-compatible endpoints) on a local machine, or Anthropic's API. No cloud requirement - the only outbound call is to the endpoint you configure.
After the first build, keep asking for changes - “make the hero full-screen”, “add a testimonials section”, “switch to a dark theme” - and the AI patches the page in place.
Upload images and the AI weaves them into the design using exact relative paths. Anything wider than 1920 px is resized automatically.
A backup is created before every generation. Preview any previous version full-screen and restore it with one click - restores are backed up too, so nothing is ever lost.
Clean, semantic, mobile-first HTML/CSS/JS on Bootstrap 5.3 - with a cookie notice and privacy statement included, plus Open Graph tags injected automatically for proper link previews.
Every project downloads as a ready-to-host autonome.zip. Drop it on any static host - no server, no dependencies.
Every user gets their own projects and files. Admins manage the AI configuration centrally and can promote or demote roles.
Long builds run in a detached background worker. Leave the page or close the browser - the job continues, you can re-attach when you return, and optionally get an email when it finishes.
A single folder of files served by the ASPPY runtime. No npm, no build step, no external tracking - and a broken AI response can never overwrite your site.
From prompt to page in six moves. The agent may call tools and reason along the way - but it always ends by writing exactly one fenced html block, which the server validates before anything is saved.
A running generation can be cancelled - any partial result is discarded and nothing is written. Navigate away mid-build and the app re-attaches to the still-running job when you come back.
Type your request in the chat - optionally after uploading photos. Enter sends; Shift+Enter inserts a newline.
If the server supports it, a detached worker runs the build in the background; otherwise it runs synchronously. The UI polls for status every few seconds and shows “AI is building your website”.
Your request, the project name, the current index.html (so it improves rather than rebuilds), your image paths, the last 5 prompts as context - and Wikipedia tools if enabled. It may call tools, reason, then write one fenced html block.
The answer must be a complete HTML document. Long dashes are replaced, image paths made relative, internal block-ID markers stripped and re-numbered, a responsive safety net (viewport meta + overflow guards) and Open Graph tags injected - then the previous version is backed up to backups/<timestamp>/.
Your fresh index.html lands and the preview refreshes. Keep refining in plain language, or roll back via History.
A broken or incomplete page triggers up to 3 retries with adaptive “repair” hints; otherwise your site is left untouched and the error explained. Truncated answers (finish_reason = length) come with a hint to raise max_tokens.
All navigation happens through URL fragments (#/…). The public landing page explains Autonome; everything else lives behind a session - the dashboard, the workspace where you talk to the agent, and the admin panel.
The workspace (#/project/<id>) is a split screen: live preview on the left, the AI Builder chat on the right. Uploaded photos appear as a thumbnail strip above the preview - hover one to delete it.
“Suggest improvement” opens starter prompts on a fresh project, or grouped ideas (tone, theme, layout, colors, content structure) on an existing one. Tick “Email me when the generation is complete” for a notification email per build.
The dashboard lists every project you own, most recently updated first. Click a card to open its workspace.
Updated 2 h ago · 14 versions
Updated yesterday · 6 versions
Updated 3 d ago · 2 versions
Each project lives in a folder with a random 16-character name generated at creation - never derived from the project name or username. URLs stay stable and identities are never leaked into paths.
Existing username + password. Admins are routed to the Admin panel automatically; everyone else lands on the Dashboard.
Username (plain ASCII), a valid email, and a password of at least 6 characters. New accounts are regular users - not admins.
Enter your email; if it's registered (and SMTP is configured) a reset link - valid for 1 hour - is emailed to you. The response is intentionally neutral so addresses can't be probed.
| Field | Value |
|---|---|
| Username | admin |
| Password | admin123 |
The default account is created on first run and is widely known. Change it right away (Account → Change password).
Only administrators see the Admin tab. Three panels: LLM configuration, email settings, and user management.
| Setting | Description |
|---|---|
| LLM Endpoint URL | Base URL of your LM Studio / Ollama server, e.g. http://localhost:1234. OpenAI-compatible. |
| API Key | Needed for LM Studio remote access (and any OpenAI/Anthropic key). |
| Model Name | Click ↻ Models to fetch the list from the endpoint (OpenAI /v1/models; Anthropic returns a curated list). Or type one manually with the ✎ Edit toggle. |
| Wikipedia tools | On/off - lets the AI research topics via search_wikipedia / get_wikipedia_page while building. |
| System Prompt | Optional extra instructions prepended to every generation. |
| Extra LLM Parameters | Optional JSON, e.g. {"temperature": 0.7, "max_tokens": 8192}. |
| Internal URL | The address the app uses to reach itself for the async worker (bypasses a reverse proxy). Must match the port ASPPY listens on - e.g. http://127.0.0.1:8058 behind IIS ARR. |
http://localhost:8080.Emails are only sent for the password-reset flow and when a user ticks “Email me when the generation is complete”.
| Username | Role | Created | |
|---|---|---|---|
| admin | root@local | admin | 2026-09-16 |
| jmeier | j@local | user | 2026-09-14 |
Promote users to admin, or demote them back - one click per row.
The whole application is a handful of files you can keep in any folder - the www folder in this project is just one example. The database is created automatically on the first request. No setup step needed.
Python 3.9+ on Windows, Linux or macOS - then ASPPY from pip. The ASPPY source folder is not included with, or needed by, this app.
You'll also need an LLM endpoint (local LM Studio / Ollama, or an Anthropic key) and - optionally - an SMTP account for emails.
Point the docroot argument at whatever folder contains the app files - any name, anywhere:
On Windows, double-click start_www.bat - it frees port 8080, starts the server and opens your browser. In a source checkout the equivalent is python ASPPY/server.py …; after pip install, asppy alone is enough.
Open http://localhost:8080, sign in with the default admin, change that password, then set your LLM endpoint in the Admin panel.
Then: Dashboard → + New Project → describe your site in plain language. Other users can register their own accounts from the sign-in page.
Everything lives inside the app folder you serve with ASPPY. The UI is a React single-page app; the backend is one REST JSON API backed by SQLite.
The app is just the files above. Keep them in www, autonome or wherever - ASPPY only needs a docroot path.
Project folder names are generated at creation and never derived from the name or username - URLs stay stable and identities don't leak into paths.
Before each generation - and before each restore - the current site is copied to backups/<timestamp>/. Nothing is ever lost.
ValidateFileExtension refuses any other filename in AI output, and every document must pass a structural completeness check before it's saved.
The SPA talks to api.asp?action=… over JSON - application/x-www-form-urlencoded with a json= field, or multipart for uploads. Sessions are cookie-based; the SPA pings keepalive every 5 minutes while logged in.
Initialise/seed the database (auto-called) and manage sessions.
Credentials and email management, with single-use reset tokens.
LLM and email settings, plus model discovery from the endpoint.
Full project lifecycle - ownership is verified on every call.
Submit, detach, poll, cancel and re-attach to AI build jobs.
Multipart uploads into the project's img/, auto-resized to 1920 px.
Backups, full-screen previews, restores and the autonome.zip export.
Per-project prompt/reply logs, and role management for admins.
Health and version info.
Don't expect perfection in one shot - iterate. Small changes patch individual blocks; big changes are full redesigns you ask for explicitly.
Mention goal, audience, colors, style, tone and the content you want.
Upload your photos before the first build so the agent can design around them.
Refine in plain language across several rounds - that's what the agent is for.
The pick-list surfaces useful improvement patterns you might not think of.
Ask for a full redesign explicitly; small tweaks are handled by patching blocks in place.
Write the prompt in your own language - the generated site comes out in that same language.
If you see “cut off by the token limit”, bump max_tokens in the LLM parameters.
Every generated image carries an onerror placeholder, so a broken link never ruins the page.
Autonome makes outbound requests only to the LLM endpoint you configure and (optionally) to the public Wikipedia API. Everything else stays on your machine - here's how it's protected.
The admin / admin123 account is created on first run and is widely known. Rotate it right after your first login.
Passwords are stored as bcrypt hashes (ASPPY crypto). Reset tokens are single-use and expire after 1 hour.
The API is session-authenticated; project operations verify ownership (user_id), backups and filenames are checked against path traversal, and random folder names keep usernames out of URLs.
Your configured LLM endpoint, and optionally the public Wikipedia API. No telemetry, no analytics, no phone-home.
Only index.html is ever written by the model. Any other filename is refused, and every generated document must pass a structural completeness check before it's saved.
One folder of files, one command to start, and a conversation that ends in a finished site - hosted wherever you like, owned entirely by you.