tutorial

Unlocking the Codex harness: how we built the App Server

For AI workflow builders, this offers a proven pattern for embedding a coding agent into custom applications, enabling tighter integration and more controlled agent interactions in development pipelines.

OpenAI Blog·February 4, 2026·1 min readtutorial

tutorialUnlocking the Codex harness: how we built the App Server

openai.com

What happened

OpenAI has released a technical blog post detailing the architecture and implementation of the Codex App Server, a bidirectional JSON-RPC API designed to embed the Codex coding agent into custom applications. The server supports streaming progress updates, tool use, approval workflows, and diff generation, enabling developers to integrate Codex's capabilities into their own tools and environments. The post walks through the design choices and implementation steps, offering a practical guide for building similar agent-serving infrastructure. For developers creating AI-powered coding assistants or workflow tools, this provides a reference for handling real-time communication, security, and state management when exposing an agent via an API. The focus is on technical depth rather than product announcements, making it a resource for those building their own agent systems.

Key takeaways

OpenAI published a technical blog post on building the Codex App Server, an API to embed the Codex agent.
The server uses a bidirectional JSON-RPC protocol for streaming progress, tool calls, authorization, and diff output.
The post covers architectural decisions for real-time updates, safety approvals, and code diff generation.
It serves as a tutorial for developers wanting to create their own agent-serving infrastructure.
The implementation is open-sourced to facilitate community adoption and customization.