OpenAI Codex integrates with Amazon Bedrock for enterprise model inference

Amazon Web Services has published guidance for running OpenAI Codex with Amazon Bedrock as the inference backend for supported OpenAI models. The post outlines three rollout paths that let platform teams keep existing AWS identity, security, and observability practices while giving developers access to Codex on their workstations.

Three rollout paths for Codex on Bedrock

The simplest path is Native AWS Access, which uses IAM Identity Center for workforce sign-in and temporary credentials. Developers sign in through their existing identity provider, and Codex calls Amazon Bedrock Runtime directly. The second path, AgentCore Gateway, provides AWS managed gateway infrastructure for centralized routing and Amazon Bedrock Guardrails. The third path, an LLM Gateway, fits organizations that already operate a gateway or want to enforce controls in the request path; a custom integration is the advanced version of this option. Each path maps different control needs to an owning team, and the guidance recommends starting with the least complex path that meets the requirement.

Pilot validation and identity isolation

The guidance begins with a small, least-privilege pilot. Prerequisites include adding pilot users to a dedicated group in the workforce identity system, provisioning to IAM Identity Center, selecting an Amazon Bedrock Region and an available OpenAI model such as GPT-5.6 Sol, and granting only the required model inference permissions. The pilot validates AWS access with a direct Bedrock Runtime call, configures the Codex provider to use the approved model and profile, and runs a safe read-only task in a non-production repository. A two-user identity isolation check ensures that separate pilot users produce distinct assumed-role sessions, audit records, and telemetry while sharing the same approved Codex and Bedrock path.

Observability, cost allocation, and governance

Observability uses the AWS Distro for OpenTelemetry collector and a managed CloudWatch dashboard to capture Codex client metrics such as token usage, turns, API requests, and tool calls. Amazon Bedrock service metrics, AWS identity records, and the Cost and Usage Report provide service-side activity, attribution, and billed usage. The guidance stresses correlating these sources for the deployed path and not presenting token volume as a cost report. Control tests include alarms for detection, IAM and managed client controls for access boundaries, and gateway policy for request-path enforcement such as hard budgets or rate limits.

Phased expansion and configuration boundaries

After the infrastructure path works, the rollout standardizes developer experience through organization-owned plugins, hooks, and approved Model Context Protocol servers. Plugins distribute shared skills and MCP integrations with versioning and review. Hooks enforce policy commands at PreToolUse, PostToolUse, or SessionStart events. MCP servers create separate trust boundaries and require limited access, approved credential management, and user confirmation for external changes. The administrator launch checklist records the selected path, owning team, Region, model, and pilot permissions, and requires a non-production test covering sign-in, model availability, and a safe Codex task. Cleanup removes Codex configuration, IAM assignments, gateway access, telemetry collectors, and pilot dashboards, verifying that new invocations are denied.

Sources: AWS Machine Learning Blog