Skip to main content
This runbook connects Codex CLI to GPT models on Amazon Bedrock through Bifrost. The current Bedrock-hosted GPT models expose the Responses API through Bedrock, so this setup uses Bifrost’s bedrock provider.
Codex custom providers can technically send other model families through a compatible Responses endpoint. This runbook deliberately uses GPT models because they are the natural fit for Codex and Bedrock exposes their Responses API directly.

​
Choose the model mappings

In this guide, a deployment name is the stable alias that you configure on a Bifrost provider key. Codex sends that name to Bifrost; Bifrost replaces it with the corresponding Bedrock model ID before inference. Sources: GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, and GPT-5.4. The recommended deployment names are Bifrost deployment mappings, not AWS-defined identifiers. You may choose another name, but the value selected in Codex must match the Bifrost deployment name exactly and should retain the bedrock/ provider prefix.

​
1. Configure the Bedrock provider in Bifrost

In Bifrost, go to Models > Model Providers > AWS Bedrock, then add or edit the provider key that Codex will use.
Bifrost Deployments table showing deployment names mapped to provider model IDs

Map each Bifrost deployment name to its Bedrock model ID. Add only the models that you intend to expose to Codex.

Deployment mappings do not automatically expand a restricted provider key’s model allowlist. If models is not *, include every deployment name in models exactly as it appears in deployments.
See AWS Bedrock for provider behavior.

​
2. Configure the virtual key

Create or edit the virtual key used by Codex:
  1. Allow the bedrock provider.
  2. Confirm the virtual key is active and has sufficient budget and rate limits.

​
3. Setup harness

Bifrost Edge provides the more seamless setup: it routes Codex CLI traffic through Bifrost at the machine level, so you do not need to configure a custom provider, gateway URL, or virtual key in Codex.
  1. Install or deploy Bifrost Edge by following Deploy with MDM. If Edge is already installed, skip this step.
  2. Complete the one-time setup approval and sign in through your browser.
  3. Open Edge from the menu bar or system tray, select the virtual key configured in the previous step, and confirm that Edge is connected.
  4. Confirm that Codex CLI is allowed by your organization’s AI app policy.
  5. Install Codex using the official Codex CLI guide, then start it normally. No changes to ~/.codex/config.toml or BIFROST_API_KEY are required.
Edge now routes Codex CLI requests through Bifrost in the background. For more detail about the user experience, see How Edge works.

​
4. Test Codex

With Edge, start Codex normally:
Without Edge, start Codex with the configured default, or select the model explicitly:
Inside Codex, run /status. Without Edge, confirm the provider is bifrost_bedrock and the model is bedrock/gpt-5.5. With Edge, no custom Bifrost provider appears in Codex; confirm that Edge is connected instead. Then send:
Open Logs in Bifrost and confirm the request used bedrock and resolved the expected deployment to its Bedrock model ID.

​
Model listing and the Codex model picker

The behavior in this section applies when Codex is configured directly without Edge. Bifrost’s model endpoint and Codex’s interactive picker are separate behaviors:
  • GET /openai/v1/models is the model list exposed by the Bifrost virtual key.
  • The current Codex slash command is /model.
  • Codex’s /model picker is primarily populated from the Codex model catalog. A successful Bifrost model-list response does not guarantee that every Bifrost deployment will appear in the picker.
  • Passing bedrock/gpt-5.5 explicitly is the deterministic setup even when the picker does not list it.
Do not select an unrelated bundled model while model_provider = "bifrost_bedrock" is active. Codex sends the selected model through the same Bifrost provider. If Bifrost cannot map that model, the request can fail with could not auto resolve a provider.

​
Show Bifrost deployments in the model picker

To make a Bifrost deployment appear in /model, copy a complete, compatible model entry from ~/.codex/models_cache.json into a local catalog such as ~/.codex/bifrost_catalog.json, then change its slug to the provider-qualified deployment name—for example, bedrock/gpt-5.5. Reference that catalog from the user-level ~/.codex/config.toml:
Restart Codex after saving both files. The catalog schema and capability metadata are Codex-version-specific, so preserve the copied entry’s remaining fields. For the complete entry shape and field-by-field guidance, see Listing non-OpenAI models in the model picker.

​
Switching models

Switch explicitly to another configured deployment:
If the installed Codex version opens the picker instead of accepting the argument, restart with: