AWS Bedrock

Authentication

Every Mantle route accepts a Bedrock API key, which is sent as a bearer token:

1

$platform = Factory::createPlatform(apiKey: 'ABSK...', region: 'us-east-1');

When no API key is given, requests are signed with AWS SigV4 using the standard credential chain, so the usual AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY environment variables, instance roles or profiles apply. A custom AsyncAws\Core\Credentials\CredentialProvider can be passed as well:

1

$platform = Factory::createPlatform(region: 'us-east-1');

Bedrock API keys are scoped to the region they were created in, so the key and the region argument have to agree.

Chat Completions

The default route reuses the wire protocol of the symfony/ai-generic-platform bridge, which it requires:

1
2
3
4
5
6
7
8
9
10
11
12
13

use Symfony\AI\Platform\Bridge\Bedrock\Mantle\Factory;
use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;

$platform = Factory::createPlatform(apiKey: 'ABSK...', region: 'us-east-1');

$messages = new MessageBag(
    Message::forSystem('You are a pirate and you write funny.'),
    Message::ofUser('What is the Symfony framework?'),
);
$result = $platform->invoke('openai.gpt-oss-120b', $messages);

echo $result->asText();

Responses

api: 'responses' reuses the wire protocol of the symfony/ai-open-responses-platform bridge. That package is only required when using this route:

1
2
3

$platform = Factory::createPlatform(apiKey: 'ABSK...', region: 'us-east-1', api: 'responses');

$result = $platform->invoke('google.gemma-4-31b', $messages);

The Gemma family reports a reasoning trace, which the endpoint only emits when the request asks for it through the reasoning option.

Anthropic Messages

api: 'messages' serves Claude through the Anthropic-compatible Messages API. It reuses the contract and result converter of the symfony/ai-anthropic-platform bridge, which the Bedrock bridge already requires, so this route needs no optional package:

1
2
3
4
5
6
7
8
9
10
11
12

use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;

$platform = Factory::createPlatform(apiKey: 'ABSK...', region: 'us-east-1', api: 'messages');

$messages = new MessageBag(
    Message::forSystem('You are a concise and helpful assistant.'),
    Message::ofUser('Explain the Symfony framework in one sentence.'),
);
$result = $platform->invoke('anthropic.claude-haiku-4-5', $messages);

echo $result->asText();

Prompt caching is on by default: the cacheRetention argument accepts none, short (default) or long and controls the cache_control markers injected into the system prompt, the messages and the tool definitions. A workspace can be passed to scope requests to a Bedrock Mantle workspace, which is sent as the anthropic-workspace header:

1
2
3
4
5
6
7

$platform = Factory::createPlatform(
    apiKey: 'ABSK...',
    region: 'us-east-1',
    api: 'messages',
    cacheRetention: 'long',
    workspace: 'proj_example',
);

Both arguments are rejected with an InvalidArgumentException on the other two routes, rather than being silently ignored.

Unlike the other two routes, this one does not support structured output: AWS rejects output_config.format on the Mantle Messages path, so the bridge fails fast with an InvalidArgumentException rather than letting the request 400.

Model Catalogs

Each route ships a model catalog with the models verified against the endpoint. Additional models can be registered by passing them to the catalog constructor, see Working with Model Catalogs.

Mind that Mantle partitions its OpenAI-compatible surface across two path prefixes, and each one rejects the other's models with a 400: the gpt-oss and qwen families answer on /v1/..., the Gemma family on /openai/v1/.... The catalogs follow that split - Gemma is registered on the Responses route, the open-weight models on Chat Completions - so the defaults need no further thought. Registering a model of the other family means passing the matching path as well:

1
2
3
4
5
6
7
8
9

$platform = Factory::createPlatform(
    apiKey: 'ABSK...',
    region: 'us-east-1',
    modelCatalog: new ModelCatalog(['google.gemma-4-31b' => [
        'class' => CompletionsModel::class,
        'capabilities' => [Capability::INPUT_MESSAGES, Capability::OUTPUT_TEXT],
    ]]),
    path: '/openai/v1/chat/completions',
);

Code Examples

• Bedrock Mantle Chat
• Bedrock Mantle Chat with SigV4
• Bedrock Mantle Streaming
• Bedrock Mantle Tool Calling
• Bedrock Mantle Responses
• Bedrock Mantle Responses Streaming
• Bedrock Mantle Anthropic Messages
• Bedrock Mantle Anthropic Messages Streaming
• Bedrock Mantle Anthropic Messages Tool Calling