Bedrock Model Access
Most LLMs you'll use are hosted on Amazon Bedrock — Claude, Amazon Nova, Titan, Llama, etc. Bedrock requires per-account, per-model access approval. Until that's granted, your agents will get AccessDenied errors when they try to call a model.
This page is your one-time setup for that.
When to do this
Right after the deployment finishes, before you let anyone start chatting.
Steps for each model you want
The example below uses Claude Sonnet 4.6 — repeat for any other model you plan to use.
1. Open Bedrock in the AWS console
AWS Console → Amazon Bedrock → Model catalog → pick the model (e.g., Claude Sonnet 4.6).
2. Submit the use-case form
Click the Request model access flow. AWS asks for:
- Use-case description.
- Industry.
- Geography.
- Whether output is shown to end users.
- Whether the application is for internal or external use.
Fill the form honestly. Submit.
Approval typically takes about 15 minutes.
3. The form doesn't show? Use the playground workaround
Sometimes the request form isn't surfaced for an account. Workaround:
- From the model page, click Open in Playground.
- Type any short message and send it.
- You'll get an
AccessDeniederror — this automatically files the access request behind the scenes. - Wait ~15 minutes.
- Try again. The model should now work.
4. Confirm
Open the Admin Dashboard → create a test agent on the new model → send a message. If you get a reply, access is good.
If access stays denied for hours
This usually means an account-level Bedrock restriction, not a per-model issue. Some AWS accounts (newer ones, or accounts in certain regions / jurisdictions) need extra verification before any Bedrock model can be enabled.
The error message in that case usually links you to the AWS Support flow. Open a support case:
| Field | Value |
|---|---|
| Issue type | Customer service |
| Service | Account management |
| Category | Bedrock allowlisting |
Explain that you're trying to use Bedrock for an internal AI agents platform. AWS reviews and (if approved) flips the account-level flag. After that, your earlier per-model requests should work.
Models you'll likely want
The platform's catalog (gateway/llm/models.yaml) lists every Bedrock model the gateway knows how to call — Anthropic Claude family, Amazon Nova family (Lite, Pro, Premier, Sonic, Sonic 2), Meta Llama, Mistral, DeepSeek, Qwen, Kimi, and several Bedrock cross-region inference profiles (us.*, global.*). For a typical first deployment, request:
- A Claude model — Sonnet 4.x is a good text default for tools and Knowledge Bases.
- Amazon Nova Sonic (and Nova Sonic 2 when available) — for voice agents.
- Amazon Titan Text Embeddings V2 — Knowledge Base embeddings (the platform's default).
You can request access for additional models any time — the request is per-model, so you don't have to do them all at once.
Bedrock Marketplace (third-party models)
Several third-party models on Bedrock — Mistral, Qwen, DeepSeek, Kimi, and others — are sold through AWS Marketplace and require a separate subscription in addition to per-model access. If you only complete the per-model "Request model access" flow for these, they'll still fail at runtime with permission errors. From the model's page in Bedrock, follow the Subscribe link to AWS Marketplace, accept the offer, and then re-test from the Admin Dashboard.
Inference profiles
Many newer Bedrock models are exposed only via cross-region inference profiles (model IDs prefixed us. or global.). The catalog ships these as separate entries — request access against the underlying model in the same way you would for the regional one. If a model in the dashboard dropdown reports an access error after approval, double-check that you granted access to the specific profile the catalog references.
What's next
- Authentication — pick your sign-in flow.
- Deployment Management Dashboard — your control panel.