AWS Bedrock¶

For buyers already on AWS who'd rather procure inference through a Bedrock line-item than onboard another vendor. Works for Anthropic Claude, Amazon Titan, Meta Llama, Mistral, and Cohere models — whichever ones your account has been granted access to.

Koji calls Bedrock through the Converse API (POST /model/{modelId}/converse) via SigV4-signed requests. One adapter covers every vendor-qualified model ID, so switching between Claude and Titan is just a model-string change.

1. Enable Bedrock + request model access¶

This is the step that trips up new users. Bedrock is an opt-in service, and the models you actually want (Claude, Llama) require per-account activation.

In the AWS console, open Bedrock in the region you plan to use (see region routing below).
Model access → Manage model access.
Tick the models you want. Anthropic and third-party models take a few minutes to a few hours to approve. Amazon's own models (Titan, Nova) are instant.
Wait for Access granted — calls will 403 until this clears.

2. Create an IAM user with minimal permissions¶

Don't use root credentials. Don't use Action: "bedrock:*" or Resource: "*". The minimum policy Koji needs:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:Converse"
      ],
      "Resource": [
        "arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-3-5-sonnet-20241022-v2:0",
        "arn:aws:bedrock:us-east-1::foundation-model/amazon.titan-text-premier-v1:0"
      ]
    }
  ]
}

Add one ARN per model you'll use. Scoping to specific ARNs (not *) means a leaked key can't be used to run up a bill on models you didn't intend to expose.

Attach the policy to a new IAM user, then create an access key for that user. Store the key ID + secret somewhere you won't lose them — AWS only shows the secret once.

Temporary credentials¶

For production deployments, prefer short-lived STS credentials over long-lived access keys: assume a role, pass the three values (AccessKeyId, SecretAccessKey, SessionToken) into Koji, and rotate on a schedule. Koji accepts all three and passes the SessionToken through to SigV4 signing.

3. Enter the credentials in Koji¶

Settings → Model Providers → Add provider.

Field	Value
Name	e.g. `bedrock-useast1`
Provider	`AWS Bedrock`
AWS region	`us-east-1` (or whichever region — see below)
Access key ID	`AKIA...`
Secret access key	The secret paired with the key ID
Session token	Only if using STS temporary credentials; otherwise leave blank
Default model	The full Bedrock model ID — see cheat sheet below

4. Model ID cheat sheet¶

Bedrock model IDs are vendor-qualified and include a version suffix. Partial list — check Bedrock → Foundation models in your region for the authoritative list:

Family	Example model ID
Anthropic Claude Sonnet 3.5	`anthropic.claude-3-5-sonnet-20241022-v2:0`
Anthropic Claude Haiku 3.5	`anthropic.claude-3-5-haiku-20241022-v1:0`
Amazon Titan Text Premier	`amazon.titan-text-premier-v1:0`
Amazon Nova Pro	`amazon.nova-pro-v1:0`
Meta Llama 3.1 70B Instruct	`meta.llama3-1-70b-instruct-v1:0`
Mistral Large	`mistral.mistral-large-2407-v1:0`
Cohere Command R+	`cohere.command-r-plus-v1:0`

Model IDs are case-sensitive. Typos return 400 (not 404), so watch for them in extraction traces.

Region routing¶

The region you configure on the provider determines which Bedrock data-plane Koji calls — i.e. bedrock-runtime.us-east-1.amazonaws.com vs bedrock-runtime.eu-central-1.amazonaws.com. Models are not globally available; a given model ID only works in regions AWS has enabled it in.

Rough guidance as of 2026-Q2 (verify on the AWS regional availability page before committing):

us-east-1 (N. Virginia) — broadest catalog, including Anthropic's latest Claude snapshots and Amazon Nova.
us-west-2 (Oregon) — second-broadest, often has Claude and Llama in parallel with us-east-1.
eu-central-1 (Frankfurt), eu-west-3 (Paris) — for EU residency; narrower catalog, sometimes one model version behind.
ap-northeast-1 (Tokyo), ap-southeast-2 (Sydney) — Claude available, Titan available, Llama spotty.

If you need the same logical "claude-sonnet-3.5" model in multiple regions, you typically need one Koji provider per region — the model IDs are the same, but the IAM ARN's region segment differs.

Cost notes¶

Bedrock bills at Amazon's per-model rates, which are typically 5–20% higher than calling the upstream vendor directly (OpenAI, Anthropic) for the equivalent model. The premium is the price of single-cloud procurement — one bill, one DPA, one set of credentials. If margins matter more than procurement simplicity, use the direct vendor path.

Check aws.amazon.com/bedrock/pricing for current per-token rates per model per region.

Known limitation: no tool use yet¶

The Bedrock Converse API supports tool use via a separate toolConfig shape (tool specs + tool-use / tool-result content blocks), distinct from both OpenAI's tools array and Anthropic's native tool_use blocks. Koji's Bedrock adapter does not wire tool use yet — schemas that rely on tool-call-style field extraction will raise NotImplementedError on a Bedrock endpoint.

If you need tool use against Bedrock-hosted Claude today, the workaround is to call Bedrock through an OpenAI-compatible proxy (LiteLLM, Portkey, AWS Bedrock Access Gateway) and configure Koji with provider: OpenAI pointing at the proxy.

Native Bedrock tool-use is tracked on the internal roadmap.