How do you build a plugin or app with AI inside, e.g. integrating the Claude API into a WordPress plugin?

Question

Accepted Answer

You build AI into an app by calling the model's API **server-side** and treating its output as untrusted input. For Claude, that means calling the **Messages API** from your backend (PHP, for a WordPress plugin) with the API key kept on the server — never in client-side JavaScript, where anyone could steal it.

## The server-side call

The key rule: the **API key lives on the server**. The browser talks to *your* endpoint; *your* endpoint talks to Claude.

```php
<?php
// In a WordPress plugin: run this in PHP (server-side), never expose the key to JS.
$resp = wp_remote_post( 'https://api.anthropic.com/v1/messages', [
  'headers' => [
    'x-api-key'         => getenv( 'ANTHROPIC_API_KEY' ), // from env/secret store, not code
    'anthropic-version' => '2023-06-01',
    'content-type'      => 'application/json',
  ],
  'timeout' => 30,
  'body' => wp_json_encode( [
    'model'      => 'claude-sonnet-4-5',   // pick a current model id
    'max_tokens' => 1024,                   // cap output → controls cost
    'messages'   => [
      [ 'role' => 'user', 'content' => $user_prompt ],
    ],
  ] ),
] );

if ( is_wp_error( $resp ) ) { /* handle network/timeout errors, maybe retry */ }
$data = json_decode( wp_remote_retrieve_body( $resp ), true );
$text = $data['content'][0]['text'] ?? '';   // validate before trusting/displaying it
```

## Beyond the basics

```text
- TOOL USE / FUNCTION CALLING → let the model call your functions (search, DB lookup)
- STREAMING                   → stream tokens for a responsive UI on long answers
- PROMPT CACHING              → cache a large static system prompt → cheaper, faster
- STRUCTURED OUTPUT           → ask for JSON, then parse + schema-validate it
- ERRORS & RATE LIMITS        → handle 429/5xx with backoff + retry; show a fallback
- COST                        → cap max_tokens, log token usage, set per-user limits
```

Always **validate model output before using it**: never run it as code, render it without escaping, or write it to your DB unchecked. The model is a powerful but fallible text generator, not a trusted source.

## Why it matters

Shipping AI inside a real product is mostly the boring-but-critical engineering around the call, not the call itself. Keeping the key server-side prevents theft and runaway bills; handling errors, rate limits, and cost keeps the feature reliable and affordable; and validating output before acting on it prevents the model's mistakes or an injected instruction from becoming a security hole. Get those right and the Messages API becomes a dependable building block; skip them and an impressive demo turns into a leaked key, a surprise invoice, or an exploit.