Amazon Bedrock¶

The Bedrock provider runs any model hosted on Amazon Bedrock — Amazon Nova, Claude, Llama, Mistral and the rest — with the AWS credentials you already have. Use it when your organisation standardises on AWS and wants model calls to stay inside its account, region and IAM policy.

Bedrock speaks its own wire protocol, so it cannot ride the OpenAI-compatible table. This provider uses Bedrock's Converse API only. Converse is the unified surface AWS added so one request shape covers every model, including tool use and streaming, so bedrock:us.amazon.nova-micro-v1:0 and a Claude model on Bedrock run the same code.

Getting started¶

1. Install the extra¶

pip install "tulip-agents[bedrock]"

boto3 is imported lazily, so nobody who never names a bedrock: model pays for it.

2. Use your usual AWS credentials¶

No key argument is invented. The standard boto3 chain applies: environment variables, the shared credentials file, a named profile, SSO, an instance role, or IRSA in EKS.

3. Address a model¶

from tulip.agent import Agent

agent = Agent(model="bedrock:us.amazon.nova-micro-v1:0", system_prompt="You summarise tickets.")

The part after bedrock: is the Bedrock model or inference-profile id.

4. Or configure it explicitly¶

from tulip.agent import Agent
from tulip.models.native.bedrock import BedrockModel

model = BedrockModel(
    model="us.amazon.nova-lite-v1:0",
    region="us-west-2",
    profile="analytics",
    max_tokens=2048,
)
agent = Agent(model=model)

Configuration¶

Setting	Argument	Default
Model or inference-profile id	`model`	`us.amazon.nova-lite-v1:0`
Region	`region`	the boto3 session's region
Named profile	`profile`	the default credential chain
Explicit credentials	`aws_access_key_id`, `aws_secret_access_key`, `aws_session_token`	not set
Endpoint override (VPC endpoint, local mock)	`endpoint_url`	not set
Bedrock guardrail	`guardrail_id`, `guardrail_version`	not set
Retries for throttling and transient 5xx	`max_retries`	`3`, botocore adaptive mode
Per-request timeout, seconds	`request_timeout`	`120.0`
Output cap / sampling	`max_tokens`, `temperature`, `top_p`	`4096`, `0.7`, `0.9`

What differs from other providers¶

No native structured output. Converse has no response_format. It could be emulated by forcing a single tool, but tool-choice support varies by model on Bedrock, so the provider does not claim it; the agent falls back to prompted JSON, which every Bedrock model handles.

Common gotchas¶

Symptom	Likely cause
`ImportError` naming `tulip-agents[bedrock]`	The `bedrock` extra is not installed.
`AccessDeniedException`	The IAM principal lacks `bedrock:InvokeModel` / `bedrock:InvokeModelWithResponseStream`, or model access is not enabled in the Bedrock console.
`ValidationException` about the model id	Some models need an inference-profile id (`us.…`) rather than the base model id.
`NoRegionError`	No region from `region=`, the profile, or `AWS_REGION`.

Source¶

tulip/models/native/bedrock.py