Agent-Served Site

A WordPress plugin where the website has no templates. A short markdown spec (site-spec.md) is the entire source of truth. Every page is generated on the fly by an AI agent that reacts to what the visitor does — page load, clicks, form submits, typing — and renders the HTML for the next page state. Navigation, a menu, a cart, a checkout form: all invented per-request by the model, with continuity carried across the session.

It's a proof-of-concept of a "generative UI" / LLM-served site, built as a consumer of Automattic/agents-api — a worked example of how to plug a product onto that substrate.

⚠️ Proof of concept. The agent's HTML+JS output is injected into the page by design (this is the whole point), so it is XSS-by-construction. Don't run this on a production site without an output policy. See Security.

How it works

Browser shell (iframe)                     WordPress
─────────────────────                      ─────────
load / click / submit  ──POST /chat──▶  agents-api  ──▶  agent-served-site handler
                                        agents/chat        │  site-spec.md  +  session history  +  event
                                        (canonical)        │            │
                                                           │            ▼
        rendered HTML  ◀──── reply ─────────┘     wp-ai-client → AI provider (Claude, …)
        (swapped into the iframe)

1. The shell (/?served=1) is a tiny HTML document hosting a sandboxed <iframe> and a small client script.
2. On every user event the client POSTs a JSON event ({"event":"click","action":"navigate","route":"/menu"}) to the canonical agents-api REST route POST /wp-json/agents-api/v1/chat.
3. agents-api dispatches the canonical agents/chat ability to this plugin's handler, which assembles a prompt (site spec + session history + the event), calls wp-ai-client (WordPress 7.0 core), and returns the generated HTML as the canonical reply.
4. The shell swaps the HTML into the iframe and re-wires its event listeners. The loop continues, threaded by the session_id.

What comes from where

Layer	Owns
agents-api (substrate)	The canonical agents/chat contract, the REST route, the handler-dispatch filter, session/run plumbing, and the conversation store (transcript persistence).
wp-ai-client (WP 7.0 core)	Provider/model prompt execution.
This plugin (consumer)	The site spec, prompt assembly, the chat handler, and the browser shell.

agents-api ships no AI provider and no chat runtime on purpose — the consumer brings those. This plugin is that consumer.

Session history is not reinvented here: it's persisted through agents-api's conversation store (resolved via the wp_agent_conversation_store filter — the built-in CPT store, enabled by this plugin, or a host's own store), with a transient fallback when no store is registered. See includes/class-session-store.php.

Requirements

• WordPress 7.0+ (ships wp-ai-client in core).
• Automattic/agents-api active.
• An AI provider registered with wp-ai-client and authenticated — e.g. AI Provider for Anthropic.
• PHP 8.1+.

Install

1. Copy agent-served-site/ into wp-content/plugins/.
2. Ensure agents-api and an AI provider are active and the provider has an API key.
3. Activate Agent-Served Site.
4. Visit /?served=1.

Configure

• The site itself lives in site-spec.md. Edit it to change what the site is — brand, pages, rules. No code changes needed. You can also swap it at runtime via the agent_served_site_spec filter.
• Model — define AGENT_SERVED_SITE_MODEL in wp-config.php (a bare model id or provider:model). Default: anthropic:claude-haiku-4-5-20251001.

// wp-config.php — higher fidelity, slower:
define( 'AGENT_SERVED_SITE_MODEL', 'anthropic:claude-sonnet-4-6' );

Performance

The latency of a render is almost entirely output-token generation — network and prompt-processing overhead is ~2s. A full styled page is ~1,200–3,000 output tokens. Two levers keep it snappy:

• Compact-output instruction in the system prompt (concise CSS, terse markup).

• A fast model. A full page takes roughly:

  Model	Full page
  claude-haiku-4-5-20251001 (default)	~3–13s
  claude-sonnet-4-6	~25–37s

There is no token streaming in this version, so the whole page must finish before it paints; the shell shows a "Generando…" indicator meanwhile. Streaming (paint as it generates) is the natural next step.

Security

The agent returns HTML — including <style> and <script> — which is rendered directly. That is intentional for the PoC. The content runs inside a sandboxed same-origin iframe, and the chat endpoint is opened to anonymous visitors only for this plugin's agent slug. None of this is safe for production without a strict output policy. Treat it as a demo.

License

GPL-2.0-or-later.