Configuration

Publish the package configuration, then set credentials for the providers your application uses:

php artisan vendor:publish --tag=prism-config

Settings live in config/prism.php. Keep credentials in the server environment and out of browser responses, source control and model-generated options.

Provider configuration#

Each entry under providers configures one integration. Environment names are provider-specific:

For example:

OPENAI_API_KEY=your-server-side-key
OPENAI_API_FORMAT=responses
ANTHROPIC_API_KEY=your-server-side-key

Use the provider guides for all supported keys, defaults and endpoint formats. Configuration chooses credentials and routing; withProviderOptions() chooses request-specific model options.

Request timeout#

request_timeout reads PRISM_REQUEST_TIMEOUT and defaults to 30 seconds. The shared Laravel HTTP client uses it for both its connection and overall timeout. Override these per request when needed:

use Prism\Prism\Facades\Prism;

$response = Prism::text()
    ->using('anthropic', 'claude-sonnet-5-5')
    ->withClientOptions(['timeout' => 120, 'connect_timeout' => 10])
    ->withPrompt('Explain database indexing.')
    ->asText();

These client options configure the HTTP transport, not the provider's JSON body. Check Azure streaming before relying on Laravel client settings for that provider's native v1 path. Explicit media fetches have their own behavior; the provider timeout is not a global limit for every operation in your application.

Retries#

withClientRetry() accepts the Laravel HTTP client's attempt count or delay array, optional delay, retry predicate and throw flag:

use Illuminate\Http\Client\ConnectionException;
use Prism\Prism\Facades\Prism;

$response = Prism::text()
    ->using('anthropic', 'claude-sonnet-5-5')
    ->withClientRetry(
        times: 3,
        sleepMilliseconds: 250,
        when: fn ($error): bool => $error instanceof ConnectionException,
        throw: true,
    )
    ->withPrompt('Explain database indexing.')
    ->asText();

The full signature is withClientRetry(array|int $times, Closure|int $sleepMilliseconds = 0, ?callable $when = null, bool $throw = true). A delay array supplies the backoff schedule instead of a fixed delay. Retries can repeat billable requests; choose a predicate and budget appropriate for your application. See rate limits for provider quota and retry information.

Per-request provider overrides#

The third argument to using() merges an application-controlled override with the configured provider entry:

use Prism\Prism\Facades\Prism;

// $tenantApiKey is loaded after authenticating and authorizing the tenant.
$response = Prism::text()
    ->using('openai', 'gpt-4.1', ['api_key' => $tenantApiKey])
    ->withPrompt('Explain quantum computing.')
    ->asText();

Alternatively, call usingProviderConfig($config) after using(); this resolves the provider again from its configured defaults plus that override. It does not merge with earlier per-request overrides. Keep endpoint overrides under application control because requests carry provider credentials.

Telemetry#

Telemetry is disabled by default. The configuration keys and environment variables are:

Enable content capture only for a trusted sink. Prompts, completions, tool arguments and media can contain private information. The length/item limits bound opt-in content reconstruction; they do not make captured content anonymous. See telemetry for events, null fields, capture controls and operations that do not emit events.

Prism Server#

prism_server.enabled reads PRISM_SERVER_ENABLED and defaults to false. prism_server.middleware defaults to an empty array.