Bring your own AWS Bedrock

If your team holds AWS credits, connect Ellipsis to Bedrock in your own AWS account and agent model calls bill your AWS account instead of your Ellipsis bill. This is an account-level setting: it applies to every agent and requires no YAML changes. You never hand over a key. Ellipsis assumes an IAM role you create and signs each Bedrock request with the short-lived credentials that role returns.

Connect Bedrock

Open Models in the dashboard, find Bring your own AWS Bedrock, and click Connect Bedrock. The setup wizard walks four steps:

1. Details. The ARN of an IAM role in your AWS account (you create it in the next step) and the region your Bedrock inference profiles live in.
2. Grant access. Create the role with the two policies the wizard shows. The trust policy allows one Ellipsis role (its ARN is displayed) to call sts:AssumeRole, conditioned on your External ID: a value generated for your account that ensures only your Ellipsis account can use your role. Keep it in the trust policy and never rotate it. The permission policy allows only bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on Anthropic models.
3. Models. Map each canonical model id agents request (for example claude-opus-5) to the Bedrock inference-profile id in your account. At least one mapping is required.
4. Test. Ellipsis assumes your role and makes one minimal model call on your account: the token cost is negligible, and it lands on your AWS bill, which is the point. On success the connection is enabled and new sessions route to your Bedrock. On failure you get the specific cause: the trust policy or External ID doesn't match, the role isn't permitted to invoke Bedrock, or your account hasn't been granted access to the mapped model.

A saved configuration never routes until the test passes, and editing it disables routing until you test again. Test connection re-runs the check at any time; if it fails, routing is disabled rather than left to fail mid-session. One provider routes at a time: verifying Bedrock disconnects an enabled Anthropic key.

Ellipsis stores only the role ARN and the External ID; neither is a secret, and no AWS access keys exist anywhere in the setup. The credentials from each AssumeRole are short-lived and are never stored, and every call appears in your CloudTrail under a role session name prefixed with ellipsis. There is no silent fallback: if the role is deleted or its trust policy drifts while a session is running, its model calls fail with an error naming the cause instead of quietly moving back to Ellipsis-owned model access.

To disconnect, click Remove configuration; new sessions return to Ellipsis-managed models.

Supported models

The Bedrock connection carries the Claude family. Map each model your agents use to the matching inference-profile id in your account; the us. profiles below are valid from any US region, and your account needs Bedrock model access granted for each one.

A session that requests a Claude model with no mapping fails with an error naming the missing model; it is never routed to a different model or back to Ellipsis. Models outside the Claude family (see model routing) do not route through your Bedrock connection.

Billing consequences

The token line of your Ellipsis bill drops to zero; AWS bills you for the Bedrock invocations directly, against your AWS credits. Sandbox compute and the platform fee are unchanged: the fee is computed on actual usage (tokens plus compute) whether or not the tokens run through your account.

Session cost breakdowns still show real token spend, so spend limits keep working: a session that reaches its budget stops, even though the tokens are billed to your AWS account. Cached model responses are served without invoking Bedrock, so repeated identical calls (replays, retries) cost nothing on your account.