AI Server chat (web)

A complete streaming chat UI for Software Tailor AI Server in three files and zero dependencies — no npm, no bundler, no framework.

• Streams replies token by token
• Lists the models installed on your server and lets you switch between them
• Keeps the whole conversation as context, with Stop and Clear
• Turns AI Server's error responses into plain advice instead of a spinner that never ends
• Your API key never reaches the browser

python serve.py     →     http://localhost:8800

---

Run it

Prerequisites: AI Server running, an API key from AI Server → API keys, and Python 3.8+ (no packages needed).

export AISERVER_BASE_URL="http://192.168.1.42:11436/v1"   # from AI Server's Server page
export AISERVER_API_KEY="ai-suite_..."                     # from AI Server -> API keys
python serve.py

PowerShell:

$env:AISERVER_BASE_URL = "http://192.168.1.42:11436/v1"
$env:AISERVER_API_KEY  = "ai-suite_..."
python serve.py

Then open http://localhost:8800. Use PORT=9000 python serve.py to move it.

Variable	Required	Meaning
AISERVER_BASE_URL	yes	Base URL including /v1
AISERVER_API_KEY	yes	Key from AI Server → API keys
PORT	no	Local port (default 8800)

---

Why there's a serve.py and not just an HTML file

The honest answer, because it's the first thing you'll wonder:

AI Server doesn't send CORS headers, and its OPTIONS preflight requires authentication — which browsers never send on a preflight. A page opened from file:// or hosted on a different origin is therefore blocked by the browser before the request is even made.

serve.py (≈150 lines, standard library only) fixes that by serving the page and the API from the same origin: it hosts index.html and forwards /v1/* to your server.

That turns out to be the better design anyway:

	Key in the browser	Key in serve.py
Visible in devtools / localStorage	⚠️ yes	✅ no
Leaks in a screenshot or screen-share	⚠️ easily	✅ no
Reaches page JavaScript at all	⚠️ yes	✅ no

The proxy binds loopback only — it holds your key, so it must not be reachable from the network. It is a local convenience, not a shared gateway. To give other people access, point them at AI Server itself with their own keys.

How it works

File	Role
serve.py	Serves the page; forwards /v1/* upstream with the key; streams responses through unbuffered
index.html	Markup + CSS (light/dark follows your OS)
app.js	Model discovery, SSE streaming, conversation state, error messages

Worth reading if you're writing your own client:

• Streaming is SSE. Lines of data: {json} ending with data: [DONE]. app.js reads the body as a stream and keeps partial lines between chunks — the usual bug is a client that splits on \n and loses a frame cut mid-way.
• Don't buffer the proxy. serve.py flushes each chunk; without that the UI sits silent and then dumps the entire answer at once.
• Model ids are per-installation (enginea/llama3.2/3b, cloud-anthropic/…), so the picker is filled from GET /v1/models. Local, non-reasoning models sort first — reasoning models emit <think> monologues that read as broken in a plain chat window, though you can still select them.
• Errors are explained, not swallowed: 401 (bad key), 429 + X-AISuite-Upgrade (free-tier cap for generic clients — retrying won't help), 429/503 + Retry-After (busy), 502 (proxy can't reach the server).

Troubleshooting

"serve.py could not reach AI Server." Check AISERVER_BASE_URL ends in /v1. If the server is on another machine it must serve your network (Server settings → Access → This network) and be allowed through its firewall — Network diagnostics → Test connectivity on that machine checks all of it.

"The server rejected the API key (401)." Issue a fresh key in API keys and restart serve.py. Keys apply immediately on the server side; serve.py only reads yours at startup.

Model list is empty. Install a model from AI Server's Models page.

Learn more

• AI Server API & clients
• ai-server-quickstarts — the same API in curl, Python, Node.js, C# and PowerShell

Licence

MIT — copy any of this into your own project.