Codex CLI with Amazon Bedrock: Native AWS Provider Configuration and Enterprise Deployment

codex-cliamazon-bedrockawsenterprisemodel-providersconfigurationgpt-oss

Codex CLI with Amazon Bedrock: Native AWS Provider Configuration and Enterprise Deployment

Introduction

On 28 April 2026, OpenAI and AWS jointly announced that OpenAI’s frontier models and Codex are available on Amazon Bedrock in limited preview1. For enterprises already committed to AWS spend, this removes the friction of a separate OpenAI billing relationship — Codex usage counts toward existing AWS cloud commitments2.

Codex CLI gained first-class Amazon Bedrock support in v0.124.0 (23 April 2026), shipping a built-in amazon-bedrock model provider with AWS SigV4 signing and credential-chain authentication3. This article covers the native provider configuration end-to-end: authentication, config.toml profiles, available models, current limitations, and enterprise deployment patterns.

Architecture Overview

sequenceDiagram
    participant Dev as Developer Terminal
    participant CLI as Codex CLI
    participant SigV4 as AWS SigV4 Signer
    participant Mantle as Bedrock Mantle<br/>(Responses API)
    participant Model as OpenAI Model<br/>(GPT-OSS)

    Dev->>CLI: codex --profile bedrock "task"
    CLI->>SigV4: Sign request with AWS credentials
    SigV4->>Mantle: POST /v1/responses (signed)
    Mantle->>Model: Route to GPT-OSS-120B
    Model-->>Mantle: Streaming response
    Mantle-->>CLI: Tool calls + text
    CLI-->>Dev: Execute tools, display output

The Bedrock integration uses a service called Bedrock Mantle, which exposes the OpenAI Responses API (/v1/responses) under AWS SigV4 authentication4. This is distinct from the standard bedrock-runtime service — the SigV4 service name is bedrock-mantle4.

Prerequisites

  • Codex CLI v0.124.0 or later — run codex --version to confirm
  • AWS credentials configured via environment variables, ~/.aws/credentials, or IAM role
  • Bedrock model access enabled for OpenAI models in your AWS account (limited preview — registration required)1
  • Region: us-east-1 confirmed available; check the AWS console for additional region availability

Basic Configuration

Minimal config.toml

Create or edit ~/.codex/config.toml:

model = "openai.gpt-oss-120b"
model_provider = "amazon-bedrock"
web_search = "disabled"

[model_providers.amazon-bedrock]

[model_providers.amazon-bedrock.aws]
region = "us-east-1"

Three settings are critical:

  1. model_provider = "amazon-bedrock" — activates the built-in Bedrock provider with SigV4 signing3
  2. web_search = "disabled" — Bedrock Mantle only supports function and mcp tool types; the default web_search tool triggers an error4
  3. region — determines the Bedrock Mantle endpoint

Authentication

Codex CLI uses the standard AWS credential chain. The most reliable approach for local development:

# Export credentials to environment (recommended)
eval "$(aws configure export-credentials --format env)"

# Verify credentials are available
aws sts get-caller-identity

Without explicitly exporting credentials, you may encounter: failed to load AWS credentials: the credentials provider was not properly configured4.

For named profiles:

export AWS_PROFILE=codex-bedrock
eval "$(aws configure export-credentials --format env --profile codex-bedrock)"

Profile-Based Configuration

For teams using multiple providers, define a Bedrock profile alongside your default:

# Default: direct OpenAI API
model = "gpt-5.4"
model_provider = "openai"

[profiles.bedrock]
model = "openai.gpt-oss-120b"
model_provider = "amazon-bedrock"
web_search = "disabled"
model_reasoning_effort = "medium"
sandbox_mode = "read-only"

[profiles.bedrock-fast]
model = "openai.gpt-oss-20b"
model_provider = "amazon-bedrock"
web_search = "disabled"
model_reasoning_effort = "low"

[model_providers.amazon-bedrock]

[model_providers.amazon-bedrock.aws]
region = "us-east-1"

Switch at invocation time:

codex --profile bedrock "Refactor the auth module to use middleware pattern"
codex --profile bedrock-fast "Add JSDoc to all exported functions"

Available Models

As of April 2026, two OpenAI models support the Responses API on Bedrock Mantle4:

Model ID Parameters Strengths Typical Latency
openai.gpt-oss-120b 120B Complex refactoring, architecture, multi-file changes 2–4s
openai.gpt-oss-20b 20B Quick edits, documentation, simple generation 1–2s

GPT-5.5 is announced for Bedrock but was not yet available for CLI API-key authentication at time of writing5. Other Bedrock models (Claude, Nova, Llama) do not support the /v1/responses endpoint required by Codex CLI v0.124+4.

Enterprise Deployment Patterns

IAM Policy for Codex Users

Grant least-privilege access to Bedrock Mantle:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": [
        "arn:aws:bedrock:us-east-1::foundation-model/openai.gpt-oss-120b",
        "arn:aws:bedrock:us-east-1::foundation-model/openai.gpt-oss-20b"
      ]
    }
  ]
}

CI/CD with GitHub Actions

name: Codex PR Review
on:
  pull_request:
    types: [opened, synchronize]

jobs:
  review:
    runs-on: ubuntu-latest
    permissions:
      id-token: write
      contents: read
    steps:
      - uses: aws-actions/configure-aws-credentials@v4
        with:
          role-to-arn: arn:aws:iam::123456789012:role/codex-ci
          aws-region: us-east-1

      - uses: actions/checkout@v4

      - run: npm i -g @openai/codex

      - run: |
          codex exec \
            --model openai.gpt-oss-120b \
            --model-provider amazon-bedrock \
            --approval-mode full-auto \
            "Review the diff for bugs and security issues. Output a markdown summary." \
            > review.md

For regulated environments, Bedrock supports VPC endpoints via AWS PrivateLink1. This keeps all traffic between your VPC and Bedrock off the public internet:

graph LR
    A[Developer Workstation] -->|VPN| B[Corporate VPC]
    B -->|PrivateLink| C[Bedrock Mantle Endpoint]
    C --> D[GPT-OSS Models]
    style C fill:#f9f,stroke:#333

Configure the VPC endpoint for the com.amazonaws.us-east-1.bedrock-runtime service, then set a custom base URL if needed in your managed configuration.

Managed Configuration for Teams

Distribute Bedrock settings via requirements.toml (pushed through MDM or filesystem):

[model_providers.amazon-bedrock.aws]
region = "us-east-1"

[requirements]
model_provider = "amazon-bedrock"
web_search = "disabled"

This ensures all developers in your organisation route through Bedrock without individual configuration6.

Alternative: Bedrock Access Gateway

Before v0.124.0’s native support, the community used the Bedrock Access Gateway — a Lambda function URL that translates OpenAI-compatible requests to Bedrock API calls7. This pattern remains useful if you need:

  • Access to non-OpenAI models (Claude, Nova, Llama) via Codex CLI
  • Custom request transformation or logging
  • Regions where Bedrock Mantle is not yet available
[profiles.gateway]
model = "us.amazon.nova-premier-v1:0"
model_provider = "bedrock-gw"
web_search = "disabled"

[model_providers.bedrock-gw]
name = "Bedrock Gateway"
base_url = "https://YOUR_LAMBDA.lambda-url.eu-west-1.on.aws/api/v1"
env_key = "BEDROCK_GW_API_KEY"
wire_api = "responses"

Note: this gateway approach requires maintaining the Lambda infrastructure yourself.

Known Limitations

Limitation Impact Workaround
Only GPT-OSS models support Responses API Cannot use Claude/Nova/Llama natively Use Bedrock Access Gateway
web_search must be disabled No built-in web search capability Use an MCP-based web search server
GPT-5.5 not yet available via API auth Cannot access latest frontier model Use direct OpenAI API for GPT-5.5
Limited preview regions May not be available in your preferred region Check AWS console for current availability
Chat Completions API deprecated Older Codex versions may break Upgrade to v0.124.0+

Cost Considerations

Routing through Bedrock offers several financial advantages for AWS-committed organisations:

  • Committed spend drawdown — Codex usage counts toward existing AWS commitments2
  • Consolidated billing — single invoice through AWS, no separate OpenAI account
  • AWS Savings Plans — potential applicability depending on your agreement structure

However, compare unit pricing carefully. The Bedrock markup over direct OpenAI API pricing varies by model and commitment level. For cost-sensitive workloads, profile switching lets you route expensive tasks through whichever provider offers better economics:

# Heavy architecture work via cheaper provider
codex --profile bedrock "Design the event sourcing system"

# Quick iterations via direct API (if cheaper for your plan)
codex --profile default "Fix the typo in the error message"

Troubleshooting

“The model does not support the ‘/v1/responses’ API” You specified a non-OpenAI model. Only openai.gpt-oss-120b and openai.gpt-oss-20b work with the native provider4.

“failed to load AWS credentials” Run eval "$(aws configure export-credentials --format env)" to export credentials to environment variables4.

“web_search tool type not supported” Add web_search = "disabled" to your profile or top-level config4.

Slow cold starts First request after idle may take 5–10 seconds as Bedrock provisions capacity. Subsequent requests typically complete in 2–4 seconds4.

Conclusion

The native Amazon Bedrock provider in Codex CLI v0.124+ is a pragmatic choice for AWS-invested organisations wanting consolidated billing and infrastructure controls without sacrificing Codex CLI’s agentic capabilities. The configuration is minimal — three lines in config.toml plus AWS credentials — but the current model selection is limited to GPT-OSS variants. As Bedrock Mantle adds GPT-5.5 support and expands regionally, this will likely become the default path for enterprise Codex CLI deployments on AWS.

Citations

  1. AWS. “Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents (Limited Preview).” 28 April 2026. https://aws.amazon.com/about-aws/whats-new/2026/04/bedrock-openai-models-codex-managed-agents/ ↩ ↩2 ↩3

  2. OpenAI. “OpenAI models, Codex, and Managed Agents come to AWS.” 28 April 2026. https://openai.com/index/openai-on-aws/ ↩ ↩2

  3. OpenAI. “Changelog — Codex CLI v0.124.0.” 23 April 2026. https://developers.openai.com/codex/changelog ↩ ↩2

  4. DevelopersIO. “I tried Amazon Bedrock support for Codex CLI v0.124.0.” April 2026. https://dev.classmethod.jp/en/articles/codex-cli-amazon-bedrock-mantle-responses-api/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10

  5. OpenAI. “Models — Codex.” April 2026. https://developers.openai.com/codex/models ↩

  6. OpenAI. “Managed Configuration — Codex.” https://developers.openai.com/codex/config-advanced ↩

  7. DEV Community. “Use OpenAI Codex CLI with Amazon Bedrock Models — Pay As You Go.” April 2026. https://dev.to/aws-builders/use-openai-codex-cli-with-amazon-bedrock-models-pay-as-you-go-48eb ↩