Codex on Amazon Bedrock Goes GA: Configuration, Managed Agents, and the Enterprise Multi-Cloud Pivot

Updated

codex-cliamazon-bedrockawsenterprisemulti-cloudmanaged-agentsagentcoreconfiguration

Codex on Amazon Bedrock Goes GA: Configuration, Managed Agents, and the Enterprise Multi-Cloud Pivot

On 1 June 2026, OpenAI and AWS moved Codex on Amazon Bedrock from limited preview to general availability 1. The announcement — barely five weeks after the initial preview on 28 April 2 — marks a significant shift: enterprise teams with existing AWS commitments can now run Codex through Bedrock’s infrastructure with IAM governance, CloudTrail auditing, and PrivateLink connectivity, all billed against their existing AWS spend 3.

This article covers what changed between preview and GA, the concrete config.toml setup, the Managed Agents layer built on Bedrock AgentCore, and the practical trade-offs teams should evaluate before migrating from direct OpenAI access.

What Changed at GA

The limited preview shipped with gated access, a single region, and a narrow model selection 2. The GA release expands the surface considerably:

Dimension Limited Preview (28 Apr) GA (1 Jun)
Access Application-based gating Open to all Bedrock-enabled accounts
Models openai.gpt-oss-120b, openai.gpt-oss-20b GPT-5.5, GPT-5.4, openai.gpt-oss-120b, openai.gpt-oss-20b 4
Regions US East (Ohio) only US East (Ohio), US West (Oregon) for GPT-5.4; US East (Ohio) for GPT-5.5 4
Interfaces CLI only CLI, Desktop App, VS Code extension 3
Pricing Undisclosed Per-token, matching OpenAI first-party rates, counting toward AWS commitments 3

The pricing model deserves emphasis: there are no seat licences or per-developer commitments 4. Token consumption rolls into your existing AWS cloud spend, which means procurement teams that have already negotiated enterprise discount programmes or reserved capacity do not need a separate OpenAI contract.

Configuration: From Direct OpenAI to Bedrock

Switching a Codex CLI installation from direct OpenAI to Bedrock requires three changes in ~/.codex/config.toml:

model = "openai.gpt-5.5"
model_provider = "amazon-bedrock"

[model_providers.amazon-bedrock.aws]
region = "us-east-2"

Authentication Pathways

Codex checks credentials in priority order 5:

  1. Bedrock API key — set AWS_BEARER_TOKEN_BEDROCK in your environment or in ~/.codex/.env for desktop and IDE surfaces
  2. AWS SDK credential chain — shared config files, environment variables, SSO profiles, or federated identity via credential_process

For teams using AWS IAM Identity Center (the SSO path most enterprises prefer), the setup is straightforward:

aws configure sso --profile codex-bedrock
aws sso login --profile codex-bedrock
aws sts get-caller-identity --profile codex-bedrock

Then reference the profile in your Codex config:

[model_providers.amazon-bedrock.aws]
profile = "codex-bedrock"

Use temporary credentials refreshed through SSO rather than permanent access keys on developer laptops 6. Expired sessions are the most common cause of Codex failing to connect to Bedrock — the codex doctor command now surfaces authentication issues explicitly 7.

IAM Policy: Least Privilege

A minimal IAM policy for developers needs only two actions against the foundation model ARN 6:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": "bedrock:ListFoundationModels",
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": [
        "arn:aws:bedrock:*::foundation-model/openai.gpt-5.5"
      ]
    }
  ]
}

For pilots, AWS provides the managed policy AmazonBedrockMantleInferenceAccess; production deployments should scope by project, region, model, and owner 6.

The Two Endpoint Paths: Mantle vs Runtime

Bedrock exposes OpenAI models through two distinct API surfaces, and the choice matters for Codex 6:

flowchart LR
    A[Codex CLI] --> B{Endpoint Path}
    B -->|Mantle| C[bedrock-mantle.region.api.aws/v1]
    B -->|Runtime| D[bedrock-runtime.region.amazonaws.com]
    C --> E[Tool calling ✓<br/>Projects ✓<br/>Responses API ✓]
    D --> F[Streaming ✓<br/>Guardrails ✓<br/>Converse API ✓]

Mantle is the path Codex CLI uses — it supports tool calling, the Responses API, and the Projects abstraction for cost isolation 6. The model ID on this path is openai.gpt-oss-120b (without the version suffix).

Runtime uses the standard Bedrock API with model ID openai.gpt-oss-120b-1:0 and supports Bedrock Guardrails and the Converse API, but is not the path Codex CLI connects through 6.

A common misconfiguration is using the Runtime model ID in config.toml — Codex expects the Mantle ID 6.

Feature Availability on Bedrock

Not every Codex feature works through Bedrock. The official documentation is explicit about what is and is not supported 5:

Feature Bedrock Status
Local CLI, Desktop, IDE workflows Supported
MCP servers and connectors Supported
Fast Mode Not available (on-demand inference only)
Hosted plugin directory Not available
Cloud agents and review tools Not available
Image generation / voice transcription Not available
Web search tools Disabled — Mantle supports function and mcp tool types only 8

The web search limitation catches teams off guard. If your workflows depend on Codex’s built-in web search, add web_search = "disabled" to your config.toml to avoid runtime errors 8.

Managed Agents: The AgentCore Layer

Beyond running Codex against Bedrock-hosted models, the GA announcement introduced Bedrock Managed Agents powered by OpenAI — currently still in limited preview 9. This is a fundamentally different proposition: instead of running Codex locally and routing inference to Bedrock, Managed Agents run the entire agent loop as a service on Bedrock AgentCore 10.

flowchart TB
    subgraph "Codex CLI (Local)"
        CLI[Codex CLI] --> LM[Local Agent Loop]
        LM --> API[Bedrock Mantle API]
    end

    subgraph "Managed Agents (AgentCore)"
        MA[Agent Definition] --> RT[AgentCore Runtime]
        RT --> LOOP[Managed Agent Loop]
        LOOP --> TOOLS[Tools & Skills]
        LOOP --> MEM[Persistent Memory]
        LOOP --> ID[Per-Agent IAM Identity]
    end

The key differentiators of Managed Agents 10:

  • Persistent memory that survives session boundaries — the agent retains context across days, not just within a single conversation
  • Skills that encode procedures the agent can invoke, similar to Codex CLI’s skills system but managed by the platform
  • Per-agent IAM identities with individual roles and least-privilege enforcement
  • CloudTrail logging for every reasoning step and tool invocation

For enterprise teams building long-running operational workflows — procurement processing, onboarding automation, or cross-system orchestration — Managed Agents provide the runtime without requiring teams to manage the agent loop infrastructure themselves 10.

Enterprise Security Architecture

Running Codex through Bedrock changes the security posture significantly compared to direct OpenAI access 3:

Network Isolation

Configure PrivateLink interface endpoints for bedrock, bedrock-runtime, and bedrock-mantle services. Apply endpoint policies to restrict traffic to approved models and actions 6. With PrivateLink, no Codex inference traffic traverses the public internet.

Audit and Compliance

CloudTrail captures every Bedrock API call, providing per-developer audit trails when combined with SSO-based authentication 3. For teams in regulated industries, this addresses the common objection that AI coding agent usage is unauditable.

Guardrails

Service Control Policies (SCPs) can block disallowed regions and models at the organisational level 6. Bedrock Guardrails — available through the Runtime path — provide content safety filtering and PII redaction, though these are not directly applicable to the Mantle path that Codex CLI uses 6.

Migration Decision Framework

Not every team should switch to Bedrock. The trade-off matrix:

Factor Direct OpenAI Bedrock
Feature completeness Full (Fast Mode, cloud agents, web search, image gen) Partial (local workflows, MCP only)
Authentication API key IAM, SSO, federated identity
Billing Separate OpenAI contract Consolidated AWS spend
Network isolation Internet-facing PrivateLink available
Audit trail OpenAI dashboard CloudTrail
Model freshness Immediate access to new models Availability lags behind direct access

The strongest case for Bedrock is an enterprise that already has AWS commitments, needs CloudTrail-grade auditing, and primarily uses Codex for local coding workflows. Teams that depend on cloud agents, Fast Mode, or web search should stay on direct OpenAI access until Bedrock parity improves.

Verification and Troubleshooting

After configuration, verify the setup:

# Confirm available OpenAI models in your account
aws bedrock list-foundation-models \
  --profile codex-bedrock \
  --region us-east-1 \
  --query "modelSummaries[?providerName=='OpenAI'].[modelId,modelName,modelLifecycle.status]" \
  --output table

# Check Codex provider status
codex doctor
# Or in the TUI:
# /status should show amazon-bedrock as the active provider

Common failure modes 6:

Symptom Cause Fix
Model not found Wrong model ID for endpoint path Use Mantle ID (openai.gpt-5.5), not Runtime ID
AccessDeniedException IAM, SCP, or region restriction Review IAM policy, SCPs, and regional model availability
Codex ignores Bedrock config Outdated CLI version Upgrade to v0.124.0+ and confirm model_provider = "amazon-bedrock"
SSO works in terminal, fails in Codex Different profile or expired session Re-login and verify profile matches config.toml
Web search errors Mantle does not support web search tools Set web_search = "disabled" in config.toml

What This Means for the Multi-Cloud Story

The Bedrock GA follows the end of Azure exclusivity announced in late April 2026 11. OpenAI models are now available through Azure OpenAI Service, Amazon Bedrock, and directly — three distinct consumption paths with different governance, billing, and feature profiles.

For Codex CLI teams, this creates a practical multi-cloud deployment pattern: different developers or CI/CD environments can point at different providers through profile-based config.toml configurations, while the agent behaviour remains identical. The --profile flag introduced in v0.134.0 makes this operationally clean — a developer can switch between bedrock-prod and openai-direct profiles without reconfiguring their entire environment 12.

⚠️ Regional availability remains limited. GPT-5.5 is currently available only in US East (Ohio), which may not satisfy data residency requirements for teams operating in the EU or Asia-Pacific.

Citations

  1. AWS and OpenAI announce expanded partnership to bring OpenAI models, Codex, and Managed Agents to Amazon Bedrock — Amazon, June 2026 ↩

  2. Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents (Limited Preview) — AWS What’s New, April 2026 ↩ ↩2

  3. OpenAI models, Codex, and Managed Agents come to AWS — OpenAI, 2026 ↩ ↩2 ↩3 ↩4 ↩5

  4. Get started with OpenAI GPT-5.5, GPT-5.4 models, and Codex on Amazon Bedrock — AWS Blog, June 2026 ↩ ↩2 ↩3

  5. Use Codex with Amazon Bedrock — OpenAI Developer Documentation, 2026 ↩ ↩2

  6. OpenAI on Amazon Bedrock: Codex, GPT-5.5, Managed Agents — Setup Guide — Elevata, 2026 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11

  7. Codex CLI v0.135 release guide — Codex Knowledge Base, May 2026 ↩

  8. Amazon Bedrock + OpenAI Models, Codex & Managed Agents (2026 Guide) — FactualMinds, 2026 ↩ ↩2

  9. Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents (Limited Preview) — AWS What’s New, April 2026 ↩

  10. Get to your first working agent in minutes: Announcing new features in Amazon Bedrock AgentCore — AWS Machine Learning Blog, 2026 ↩ ↩2 ↩3

  11. End of Azure Exclusivity: Multi-Cloud Codex and Enterprise Deployment — Codex Knowledge Base, April 2026 ↩

  12. Codex CLI Changelog — OpenAI Developer Documentation, 2026 ↩