README

Official Google Cloud Agent Platform provider for the framework-agnostic PHP AI SDK. It uses the OpenAI-compatible endpoint for text and transcription, and native publisher model endpoints for embeddings, image, and speech generation.

Installation

composer require aisdk/google-agent-platform

Basic Usage

use AiSdk\Generate;
use AiSdk\GoogleAgentPlatform;

$result = Generate::text()
    ->model(GoogleAgentPlatform::model('google/gemini-2.5-flash'))
    ->prompt('Explain closures in PHP.')
    ->run();

echo $result->text;

Embeddings

$embedding = Generate::embedding(['First document to index', 'Second document to index'])
    ->model(GoogleAgentPlatform::model('gemini-embedding-001'))
    ->dimensions(768)
    ->providerOptions('google-agent-platform', [
        'task_type' => 'RETRIEVAL_DOCUMENT',
        'autoTruncate' => false,
    ])
    ->run();

$vector = $embedding->output->vector;

The package sends one native publisher-model request per input, which supports the documented single-input limit of gemini-embedding-001. Provider options use Google's documented field names: task_type, title, autoTruncate, and outputDimensionality.

Publisher and routed model IDs pass through unchanged and do not need to be registered. This package does not ship a model inventory; the SDK performs internal adapter validation before Google Cloud validates support for the selected model, project, and location.

Image and speech generation

$image = Generate::image('A clean product photograph')
    ->model(GoogleAgentPlatform::model('google/gemini-3.1-flash-image'))
    ->aspectRatio('16:9')
    ->run();

$speech = Generate::speech('Welcome to the application.')
    ->model(GoogleAgentPlatform::model('google/gemini-3.1-flash-tts-preview'))
    ->voice('Kore')
    ->run();

Native embedding, image, and speech generation require project; a custom OpenAI-compatible baseUrl alone is not enough to construct the publisher model endpoint.

Transcription

use AiSdk\Content;
use AiSdk\Generate;
use AiSdk\GoogleAgentPlatform;

$result = Generate::transcription(Content::audio(__DIR__.'/meeting.mp3'))
    ->model(GoogleAgentPlatform::model('google/gemini-2.5-flash'))
    ->run();

echo $result->output->text;

Transcription uses the routed multimodal model endpoint, so it follows the same authentication and model-id rules as text generation.

Live voice sessions

Install the optional transport package for a ready-made WebSocket connection:

composer require aisdk/transport

use AiSdk\GoogleAgentPlatform;
use AiSdk\Live;
use AiSdk\Live\AudioDelta;
use AiSdk\Live\TranscriptDelta;
use AiSdk\Transport;

$session = Live::voice()
    ->model(GoogleAgentPlatform::model('gemini-live-2.5-flash-native-audio'))
    ->instructions('You are a concise customer-support agent.')
    ->voice('Kore')
    ->language('en-US')
    ->connect(Transport::auto());

// Send 16 kHz mono PCM chunks while reading events concurrently.
$session->sendAudio($pcmBytes);

foreach ($session->events() as $event) {
    if ($event instanceof AudioDelta) {
        playAudio($event->bytes);
    }

    if ($event instanceof TranscriptDelta) {
        echo $event->delta;
    }
}

Core automatically runs registered tools that have handlers and sends the provider's required function response. Tool calls without a matching handler are emitted for manual handling.

Provider-only setup fields can be merged without adding them to core. For example, a previously received session-resumption handle can be supplied with:

->providerOptions('google-agent-platform', [
    'raw' => ['session_resumption' => ['handle' => $resumeHandle]],
])

Resumption updates and unknown native messages are preserved as ProviderEvent values.

Agent Platform Live sessions can emit input and output transcripts, but this package does not claim dedicated Live::transcribe() or Live::translate() implementations. Google Cloud Speech-to-Text streaming is a separate gRPC service, and Gemini Developer API Live Translate's translationConfig protocol is not documented for the Agent Platform endpoint.

Core-only transport

The provider is fully usable without aisdk/transport. Pass any application implementation of the core transport contracts:

$session = Live::voice()
    ->model(GoogleAgentPlatform::model('gemini-live-2.5-flash-native-audio'))
    ->connect($appWebSocketTransport);

See the core custom-transport guide for a complete implementation. The connection receives the current native v1 Agent Platform WebSocket endpoint and OAuth, ADC, or API-key headers prepared by this package.

Agent Platform's native Live WebSocket streams media directly. With the default server-side activity detection, speech boundaries are detected automatically and the protocol does not document a separate audio commit event. When turn detection is disabled, commitAudio() sends the documented manual activity-end signal. clearAudio() remains unsupported because the protocol has no input buffer. It also does not expose the client-secret, WebRTC, or SIP lifecycle implemented by providers that officially support those topologies.

Streaming

foreach (Generate::text('Tell me a story.')->model(GoogleAgentPlatform::model('google/gemini-2.5-flash'))->stream()->chunks() as $chunk) {
    echo $chunk;
}

Video Generation

$result = Generate::video('A cinematic ocean scene')
    ->model(GoogleAgentPlatform::model('veo-3.1-generate-001'))
    ->aspectRatio('16:9')
    ->resolution('1920x1080')
    ->run(timeout: 600);

Veo operations use the native publisher-model endpoint and can return inline video bytes or a Cloud Storage URI.

Reasoning and multimodal input

Portable reasoning maps to Google's native thinking configuration. Effort levels become thinking_level, while token budgets become thinking_budget:

$result = Generate::text('Explain the tradeoff.')
    ->model(GoogleAgentPlatform::model('google/gemini-3.1-pro-preview'))
    ->reasoning(\AiSdk\Reasoning::effort('medium'))
    ->run();

Google-specific extra_body.google.thinking_config provider options take precedence when supplied explicitly. Text requests also accept documented image and audio inputs; audio formats are serialized using Google's MIME-style values such as audio/wav.

Configuration

Variable	Description
`GOOGLE_CLOUD_PROJECT` / `GOOGLE_VERTEX_PROJECT`	Google Cloud project id (required unless `baseUrl` set)
`GOOGLE_CLOUD_LOCATION` / `GOOGLE_VERTEX_LOCATION`	Region (defaults to `global`)
`GOOGLE_VERTEX_ACCESS_TOKEN`	Static OAuth access token
`GOOGLE_VERTEX_CREDENTIALS_PATH`	Path to a service account JSON key
`GOOGLE_APPLICATION_CREDENTIALS`	Standard ADC service-account path
`GOOGLE_VERTEX_API_KEY`	API key (express mode)

Authentication

Authentication is powered by the official google/auth library and supports the full set of Google credential sources:

Application Default Credentials (ADC) — used automatically when no explicit credential is given: GOOGLE_APPLICATION_CREDENTIALS, gcloud user credentials, GCE/GKE metadata server, and workload identity federation.
Service account — JSON key as an array or file path (OAuth token exchange handled for you).
Static OAuth access token — bring your own (e.g. gcloud auth print-access-token).
API key — express mode via the x-goog-api-key header.

// Application Default Credentials (nothing to configure)
GoogleAgentPlatform::create(['project' => 'my-project', 'location' => 'us-central1']);

// Service account file
GoogleAgentPlatform::create([
    'project' => 'my-project',
    'location' => 'us-central1',
    'credentialsPath' => '/path/to/service-account.json',
]);

// Service account array
GoogleAgentPlatform::create([
    'project' => 'my-project',
    'credentials' => $decodedServiceAccountJson,
]);

// Static access token
GoogleAgentPlatform::create(['project' => 'my-project', 'accessToken' => 'ya29....']);

Testing

composer test

The default suite uses protocol fixtures and conformance checks. Credentialed Live network verification is separate from the default test run.

Documentation

Community

Contributing
Support
For private security reports, email security@phpaisdk.com.

aisdk / google-agent-platform

Maintainers

Package info

Fund package maintenance!

Statistics

Security