AWS Bedrock¶

Autobot supports AWS Bedrock as an LLM provider via the Converse API. This gives access to Claude, Amazon Nova, and other foundation models hosted on AWS with SigV4 authentication.

Setup¶

1. Configure credentials¶

Add AWS credentials to your .env file:

AWS_ACCESS_KEY_ID=AKIA...
AWS_SECRET_ACCESS_KEY=wJalr...
AWS_REGION=us-east-1

Or use the interactive setup:

autobot setup
# Select "AWS Bedrock" as provider

2. Configure the provider¶

In config.yml:

agents:
  defaults:
    model: "bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0"

providers:
  bedrock:
    access_key_id: "${AWS_ACCESS_KEY_ID}"
    secret_access_key: "${AWS_SECRET_ACCESS_KEY}"
    region: "${AWS_REGION}"

3. Verify¶

autobot doctor
# Should show: ✓ LLM provider configured (bedrock)

Model naming¶

Models must use the bedrock/ prefix. After the prefix, use the Bedrock model ID:

# Foundation models
model: "bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0"
model: "bedrock/amazon.nova-pro-v1:0"

# Cross-region inference profiles
model: "bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0"

The bedrock/ prefix tells autobot to route to the Bedrock provider instead of the HTTP provider. It is stripped before sending to the API.

Guardrails¶

Bedrock Guardrails can filter harmful content. Both guardrail_id and guardrail_version are required — setting only guardrail_id logs a warning and disables guardrails.

providers:
  bedrock:
    access_key_id: "${AWS_ACCESS_KEY_ID}"
    secret_access_key: "${AWS_SECRET_ACCESS_KEY}"
    region: "${AWS_REGION}"
    guardrail_id: "abc123def456"
    guardrail_version: "1"

When a guardrail intervenes, the response includes finish_reason: "guardrail_intervened" and the guardrail trace is logged at INFO level.

Session tokens¶

For temporary credentials (e.g. from aws sts assume-role), add session_token:

providers:
  bedrock:
    access_key_id: "${AWS_ACCESS_KEY_ID}"
    secret_access_key: "${AWS_SECRET_ACCESS_KEY}"
    session_token: "${AWS_SESSION_TOKEN}"
    region: "${AWS_REGION}"

Configuration reference¶

Field	Required	Default	Description
`access_key_id`	Yes	—	AWS access key ID
`secret_access_key`	Yes	—	AWS secret access key
`region`	No	`us-east-1`	AWS region
`session_token`	No	—	Temporary session token (STS)
`guardrail_id`	No	—	Bedrock guardrail identifier
`guardrail_version`	No	—	Guardrail version (required if guardrail_id is set)

How it works¶

Unlike other providers that use OpenAI-compatible HTTP APIs, Bedrock uses:

AWS SigV4 authentication (no API keys)
Converse API endpoint (/model/{modelId}/converse)
Union-key content blocks ({"text": "..."} instead of {"type": "text", ...})
camelCase field names (toolUse, toolResult, inputSchema)

Autobot handles all format conversion transparently. Tools, MCP servers, plugins, and all other features work the same as with other providers.

Known limitations¶

No streaming — Responses are returned in full after the model finishes generating. The ConverseStream API is not used.
No credential refresh — SigV4 credentials are set at startup. Temporary credentials (STS session tokens) will not refresh when they expire; restart autobot to pick up new credentials.
Text-only content blocks — Image and document content blocks in messages and tool results are not converted. Only text is sent to Bedrock.
Tool choice is always auto — There is no configuration to force a specific tool or disable tool use per-request.

Troubleshooting¶

Enable debug logging to see request/response details:

LOG_LEVEL=DEBUG autobot agent -m "Hello"

Look for:

Bedrock: region=... model=... — confirms provider is active
Bedrock toolConfig: N tools — confirms tools are in the request
Response 200 (N bytes) — confirms API response
HTTP 4xx/5xx: ... — API errors with details

Common issues¶

"No LLM provider configured" — Model must start with bedrock/ prefix.

"HTTP 403: ..." — Check IAM permissions. The role needs bedrock:InvokeModel on the model ARN.

"HTTP 404: ..." — Model not available in the configured region. Try a cross-region inference profile (e.g. us.anthropic.claude-...).

guardrail_id without guardrail_version — Both are required. Check logs for the warning message.