vibephp

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

图片：VibePHP

---

VibePHP is a next-generation PHP runtime and web server that runs PHP faster* and better*.

There is no interpreter. There is no compiler. When a request comes in, your PHP source is handed to an AI that reads it, runs it in its head, makes up whatever it needs to (the database, the clock, the network, the truth), and hands back the HTTP response it reckons the code would have produced.

It is not deterministic. It is not cheap. It is not correct. It is, however, very vibe.

<?php
$posts = $db->query('SELECT * FROM posts ORDER BY published_at DESC');
// there is no $db. there is no database. there are no posts.
// the AI will invent five believable ones anyway. ✨

* debatable

Why

An engine built for 2026 and beyond. PHP-FPM is stable, FrankenPHP is fast, VibePHP brings all the rest:

• Hype
• Generics
• Async/await
• Inline Go or Rust

VibePHP has no parser. Your code is never compiled, only understood. The contract is void, and so is every constraint that ever held the language back. Generics work because nothing stops them. So does any syntax you can dream up.

class Prices
{
    private Map $store = {};

    public async function fetchAll(K ...$symbols): Map
    {
        // Inline Go mixed with PHP, who knows what this does except the AI
        return go {
            ch := make(chan V)
            for _, sym := range $symbols { go func() { ch <- await fetchPrice(sym) }() }
            for _, sym := range $symbols { $this->store[sym] = <-ch }
            return $this->store
        };
    }
}

Benchmark

We benchmarked VibePHP against industry-leading runtimes. The numbers speak for themselves.

Runtime	Latency (p50)	Cost / request
nginx + PHP-FPM	~1 ms	~$0.00000x
FrankenPHP	~1 ms	~$0.00000x
VibePHP	~7 s	~$0.0063

The Labor Illusion study (Buell & Norton, Harvard Business School) found that people value a service up to 8% more when they can see it working for them, even preferring the slower version that returned identical results. VibePHP increases the perceived value of your PHP website by 700,000% which offsets the increased cost per request. It is a win-win.

Quickstart

• Set an OPENAI_API_KEY in .env
• Create an index.php file in the vibe/ directory:

  <?php
  echo "Hello, world! The time is " . date('H:i:s');

• Run php artisan vibe
• Visit your PHP website at http://localhost:8000

How it works

VibePHP itself is implemented using PHP, Laravel, laravel/ai, and GPT/Claude.

   GET /posts/42
        │
        ▼
┌───────────────────┐     reads the file, doesn't run it
│     Laravel       │ ───────────────────────────────────────┐
│  routes URL→file  │                                        │
└───────────────────┘                                        ▼
                                              ┌──────────────────────────────┐
   HTTP response  ◄──── {status,headers,body} │      VibePhpRuntime (AI)     │
        ▲                                     │  "i am the php engine now"   │
        │                                     │  • executes in its head      │
        └──────────────────────────────────── │  • invents missing data      │
                                              │  • reads includes via a tool │
                                              └──────────────────────────────┘

1. A request arrives. The catch-all route maps the URL to a file in vibe/ (/about → about.php), falling back to index.php as a front controller.
2. The source is read, not run. Code + the full HTTP request context get handed to the VibePhpRuntime agent.
3. The AI “executes” it. It follows the control flow, fetches include/required files on demand via the ReadVibeFile tool, and improvises every missing piece — DB rows, date(), rand(), API payloads — plausibly and with full commitment.
4. It returns { status, headers, body }. Laravel sends that back as a real HTTP response. header('Location: …') becomes a 302. header('Content-Type: application/json') is honored. The whole charade holds up.

The core of the engine is implemented as a Laravel AI agent in App\Ai\Agents\VibePhpRuntime. It uses a custom tool to read files from the vibe/ directory on demand, so that includes and requires work as expected.

Demo

• Clone this repo
• Set an OPENAI_API_KEY in .env
• Install the project with composer run setup
• Start the server with php artisan vibe
• Visit http://localhost:8000

Then wander around:

Route	What “happens”
/	Home page listing 5 freshly-hallucinated blog posts
/about	An about page (uptime: a number the AI feels good about)
/posts/42	A full article that has never existed until you asked
/api/status	JSON endpoint with imagined load metrics

Refresh any of them. It’s never the same site twice.

Write your own page

Drop a .php file in vibe/ and it’s instantly “served.” Write whatever PHP you want — the more it relies on a database, filesystem, or external API that isn’t there, the better the vibe.

<?php // vibe/weather.php
header('Content-Type: application/json');
$city = $_GET['city'] ?? 'Paris';
$temp = fetchTemperatureFromSomeApiThatDoesNotExist($city);
echo json_encode(['city' => $city, 'temp_c' => $temp]);

fetchTemperatureFromSomeApiThatDoesNotExist() is not defined anywhere. It will return a believable temperature regardless. You’re welcome.

Configuration

config/vibe.php:

Key	Env	Default	Meaning
docroot	VIBE_DOCROOT	vibe/	Where your “executable” PHP lives
model	VIBE_MODEL	provider default	Which brain interprets your code

The provider (OpenAI) is set on the App\Ai\Agents\VibePhpRuntime agent via laravel/ai.

☁️ Vibe Cloud — coming soon

Running VibePHP on your own machine is fine, but it’s not web scale. Vibe Cloud is fully managed, serverless VibePHP hosting. Deploy with vibe up and focus on running your business.

Cloud Particles™

Vibe Cloud doesn’t bill you for servers, containers, or even functions. It bills you for particles.

• 1 particle = 1 MB of memory. Your app’s footprint is measured to the megabyte, in real time.
• Particles autoscale per MB. Allocate another array, spawn another particle. Free a string, the particle decays. You pay for exactly the memory you’re using, sampled continuously, billed per particle-millisecond.
• Memory is quantized. You cannot allocate half a megabyte. There is no 0.5 particle.
• Heisenberg Tier (Enterprise). On our top tier, you can know either how much memory your app is using or how much it costs. Never both at once.

Idle Cycle Reclamation

While the model is “executing” your PHP, your CPU is sitting idle for whole seconds, politely waiting on an inference API. That’s billions of wasted cycles, per request, per particle.

Vibe Cloud puts those idle cycles to work with Speculative Pre-Vibing.

While your request is out at the inference API, every spare core races ahead to pre-hallucinate the response to the NEXT request that might be received. By the time the user clicks to another page, the page is already rendered. Zero perceived latency, and the cycles were reclaimed, which goes on the sustainability report. Vibe Cloud is proud to be carbon-neutral-adjacent.