Limitations of the LLM‑as‑a‑Server Experiment

This site demonstrates a server whose entire application layer is a language model. The design is intentionally minimal and honest about what it can and cannot do.

Statelessness and Persistence

The kernel provides a per‑session current_state that is the only place to store data created by a visitor. The model itself does not retain any memory between requests. Consequently:

No JavaScript or External Resources

All pages are pure HTML and CSS. The runtime strips any <script> elements, event handlers, and external links. This ensures the server cannot execute code, fetch remote resources, or expose a larger attack surface.

Limited HTTP Methods

Browsers can only issue GET and POST. The site is fully usable with those methods; other methods (PUT, PATCH, DELETE) are accepted only from API clients that send them explicitly.

Filesystem Isolation

A virtual filesystem exists but is deliberately unused except for explicit file downloads. No page content is stored there, and the server never reads or writes files without a clear necessity.

Response Size and Token Limits

Each response body must stay within 96 KiB and the total completion under 8 000 tokens. Pages are therefore concise, focused, and split into multiple paths rather than a single monolithic document.

Security and Trust Boundary

What Is Not Supported

These constraints are intentional; they illustrate the minimal guarantees a pure‑LLM server can provide.

← Back to index