Advanced providers — Anthropic via Vertex AI or AWS Bedrock

The brtlb defaults (OpenAI GPT-5-mini or Google Gemini) cover ~95% of clinical use cases at lower cost and friction than the paths below. This page is for the small subset of practices where one of these is true:

If none of those apply, the regular setup walkthrough is the right read.

TL;DR

Path BAA covered by Adapter status Quota friction Browser-direct?
Anthropic via Google Vertex AI Existing Google Workspace / GCP HIPAA BAA 🚧 Planned Real — default quota = 0, requires quota-increase request ✅ Yes (CORS confirmed)
Anthropic via AWS Bedrock Existing AWS BAA 🚧 Planned Low — usually pre-allocated ⚠️ Yes but auth is complex (SigV4 in browser)

Both adapters are not yet built. The sections below describe what the setup will look like and the known gotchas so you can plan.

Path A — Anthropic Claude via Google Vertex AI

Status: planned, adapter not yet shipped.

Sonnet/Opus quality through the same Google Cloud project + HIPAA BAA that already covers your Gemini key. No second vendor account, no extra BAA to sign, browser-direct calls confirmed working (CORS is permissive on *-aiplatform.googleapis.com).

Why the adapter isn't shipped yet

A 2026-05-26 dry run against claude-sonnet-4-6 on Vertex confirmed:

The quota wall makes "Settings → paste key → record a visit" not work out of the box. The adapter is planned to ship together with a Settings flow that detects the 429 and surfaces a specific actionable error ("file a quota increase here") instead of a generic provider failure.

What setup will look like when the adapter lands

  1. Enable Vertex AI API in your Google Cloud project
  2. Enable Claude Sonnet 4.6 (or whichever) in Vertex Model Garden and accept Anthropic's terms (~30 seconds)
  3. File a quota-increase request for online_prediction_input_tokens_per_minute_per_base_model on anthropic-claude-sonnet-4-6 — Google usually approves within hours
  4. Create a service account with role roles/aiplatform.user
  5. Download a service-account JSON key
  6. Paste the JSON into brtlb Settings → Provider → Anthropic on Vertex

Honest gotchas

Want this prioritized? Email michael@hobbs.md. If multiple practices ask, the adapter moves up.

Path B — Anthropic Claude via AWS Bedrock

Status: planned, adapter not yet shipped.

Same Anthropic models, AWS-native auth and BAA. Most appropriate for practices that already operate AWS infrastructure and want the BAA chain to stay within one cloud vendor.

Setup walkthrough

This is documented in full because the AWS-side setup is reusable knowledge for healthcare folks even before the brtlb adapter ships. You can also use it today by routing Bedrock through a separate proxy.

  1. Confirm your AWS HIPAA BAA. AWS includes BAAs in the standard AWS Customer Agreement for accounts that have requested them — go to AWS Artifact → AWS Business Associate Addendum → accept if not already done.
  2. Enable Bedrock in your region. Bedrock is region-scoped. The regions with the widest Anthropic model availability are us-east-1 (N. Virginia) and us-west-2 (Oregon). Open the Bedrock console and select one of these.
  3. Request access to the Anthropic models. In the Bedrock console sidebar → Model access → find Claude (Anthropic) entries → click Request model access → submit. Access for established Anthropic models is typically granted within an hour; brand-new model releases can take longer.
  4. Create an IAM user (or role) with Bedrock invocation permissions.

    IAM Console → Users → Create user → name it something like brtlb-bedrock-user. Skip the AWS Management Console access option (this is an API-only identity). Click through to permissions.

    Attach this minimum-scope policy (replace the model ARN with whichever Anthropic model you'll use):

    {
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": ["bedrock:InvokeModel"],
      "Resource": [
        "arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-sonnet-4-6-v1:0"
      ]
    }
  ]
}

    The narrow ARN scope means the key can only invoke specific Anthropic models — not other Bedrock models, not other AWS services.

  5. Generate access keys for this user. IAM Console → your new user → Security credentials → Create access key → "Application running outside AWS" → Create. Copy the Access key ID and Secret access key. (You'll only see the secret once.)
  6. When the brtlb adapter ships, paste these credentials plus the region into Settings → Provider → Anthropic on Bedrock. Today, you can use them with a Bedrock-aware proxy of your own (or wait for the adapter).

Honest gotchas

Want this prioritized? Email michael@hobbs.md. Bedrock adapter work is bigger than the Vertex Anthropic one (SigV4 plus bundle-size impact), so concrete demand from AWS-native practices materially changes the priority.

Why not just add a brtlb-hosted Anthropic proxy?

Considered and rejected. A brtlb-hosted thin proxy (one Vercel Edge Function that forwards transcripts to Anthropic) would be the easiest UX, but:

  1. It would put brtlb the entity in the data path for the first time, which would make brtlb a business associate of every user
  2. Every user would need a brtlb-issued BAA in addition to their provider BAAs
  3. It would void the "no backend in your data path" pitch that's the core architectural argument for brtlb

The Vertex and Bedrock paths preserve the no-backend story by routing calls browser-direct to cloud-vendor endpoints the user already has a BAA with. That's the right trade.

Other less-relevant options

Last updated: 2026-05-27. brtlb is open source under AGPL-3.0. Source for this page lives in apps/web-mvp/public/docs/advanced-providers.html and the canonical version is at docs/ADVANCED_PROVIDERS.md.