Amazon Bedrock

Operator guide for turning on the Amazon Bedrock provider. Unlike every other LLM provider in AXL, Bedrock has no AXL API key: it authenticates through the AWS default credential chain, and the AWS account itself must grant model access and IAM permissions before a single request will succeed. This guide covers that AWS-side setup end to end.

For the axl-side knobs see crates/agent-core/CONFIG.md (canonical BEDROCK_* reference); for how the provider is wired see the Providers section of LLM_SYSTEM.md; for the design rationale see docs/prds/AWS_BEDROCK_PROVIDER.md.

At a glance

You need four things in place before AXL can talk to Bedrock:

1. Credentials discoverable by the AWS default chain.
2. Model access granted for each model you intend to call (Bedrock blocks all models by default).
3. IAM permissions for the Converse actions.
4. A region where your access is granted.

Then set BEDROCK_ENABLED=true and AXL constructs the provider.

Credentials

AXL loads credentials with the AWS default credential provider chain (aws-config), so it picks up whatever the AWS CLI/SDKs already use. It never sees or stores a secret. Pick whichever fits your environment:

Role assumption (role_arn + source_profile, web identity, cross-account) is handled by the chain itself; there is deliberately no AXL knob for it.

Enable model access

Bedrock denies every foundation model until you request access. In the AWS console:

Bedrock -> Model access -> Modify model access, then enable the models you want. Anthropic Claude access is usually granted quickly.

The default model AXL ships is:

us.anthropic.claude-sonnet-4-5-20250929-v1:0

so at minimum enable Claude Sonnet 4.5. All curated Bedrock ids use the us. cross-region inference-profile prefix (see section 4), so grant access in the US regions the profile spans.

Missing model access surfaces as an AWS AccessDeniedException naming the model, not a credential error - so if sts get-caller-identity works but a call is denied, check here first.

IAM permissions

The Converse and ConverseStream operations are authorized by two actions:

bedrock:InvokeModel                    # Converse (non-streaming)
bedrock:InvokeModelWithResponseStream  # ConverseStream (streaming)

For a first smoke test, the AWS-managed AmazonBedrockFullAccess policy covers everything. For a scoped policy, this is the minimum - note that an inference-profile call needs permission on both the profile ARN and the underlying foundation-model ARNs:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "BedrockConverse",
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": [
        "arn:aws:bedrock:*::foundation-model/*",
        "arn:aws:bedrock:*:<ACCOUNT_ID>:inference-profile/*"
      ]
    }
  ]
}

Tighten Resource to specific model/profile ARNs once you know which models you serve. (The foundation-model ARN has no account id - those models are AWS-owned; the inference-profile ARN carries your account id.)

Region and the us. inference-profile prefix

Set the region to one where your model access is granted:

• BEDROCK_REGION=us-east-1 (or us-west-2), or
• rely on the chain's region (AWS_REGION / the profile's region) and leave BEDROCK_REGION unset.

If no region resolves anywhere, the call fails - always have one set.

The us. prefix on the curated ids (e.g. us.anthropic.claude-sonnet-4-5-...) marks a cross-region inference profile: AWS routes the request across US regions for capacity. This is why the ids differ from the raw Anthropic-on-Bedrock model ids and why you call from a US region. To use a plain (non-profile) model id or another geography, set BEDROCK_MODEL to that id and grant access accordingly.

Turn it on in AXL

Minimum configuration (full reference in agent-core/CONFIG.md):

BEDROCK_ENABLED=true
BEDROCK_REGION=us-east-1
# BEDROCK_MODEL defaults to us.anthropic.claude-sonnet-4-5-20250929-v1:0
# LLM_PROVIDER=bedrock   # only if Bedrock should supply the default model

At init, AXL logs the resolved region, credential source, and default model id (no secret is logged), so a region/profile/model mismatch is self-diagnosable from the startup line.

Verify

Confirm the AWS side independently of AXL first - this isolates any problem:

# 1. Do credentials resolve?
aws sts get-caller-identity

# 2. Is model access granted in this region?
aws bedrock list-foundation-models --region us-east-1 --by-provider anthropic

# 3. End-to-end smoke, bypassing axl entirely:
aws bedrock-runtime converse \
  --region us-east-1 \
  --model-id us.anthropic.claude-sonnet-4-5-20250929-v1:0 \
  --messages '[{"role":"user","content":[{"text":"say hi in 3 words"}]}]'

If step 3 returns text, AXL will work - the provider is a thin wrapper over exactly that call. Then boot the server with BEDROCK_ENABLED=true and exercise it (the tester-bedrock agent in axl-config covers streaming, reasoning replay, and images).

Cost note

For plumbing checks, use a cheap model to keep it near-free: Claude 3.5 Haiku (us.anthropic.claude-3-5-haiku-20241022-v1:0) or Amazon Nova Micro (us.amazon.nova-micro-v1:0). The reasoning path (manual budget_tokens, stop_sequences under thinking) needs a reasoning-capable Claude - Sonnet 4.5, Opus 4.1, or Claude 3.7 Sonnet - so budget a few cents for those specific runs.

Troubleshooting

See also

• crates/agent-core/CONFIG.md - canonical BEDROCK_* env-var reference.
• docs/systems/LLM_SYSTEM.md - provider architecture, extended reasoning, and Bedrock Known Issues.
• docs/prds/AWS_BEDROCK_PROVIDER.md - design record and rationale.

Next

• Choose a model - selecting a Bedrock model per agent.
• LLM failover - falling back across providers during an outage.
• Model reference - registry and provider behavior.