Most developers first encounter coding agents through terminal CLIs, desktop apps, or IDE extensions. In "Codex as a Platform: Build on the Open Agent Harness", the OpenAI Developers team describes how to embed the Codex agent loop inside software built for a specific job.

Unless noted otherwise, the product descriptions and examples below summarize that OpenAI post. They are not independent performance tests.

What the harness handles

A model call does not retain a thread, stream progress, invoke tools, enforce sandbox policies, or request approval on its own. The harness handles those parts of execution.

OpenAI reports that harness choices changed one ARC-AGI-3 evaluation: retained reasoning and context compaction raised GPT-5.6 Sol's score from 13.3% to 38.3% while reducing output tokens sixfold. Those figures are OpenAI's reported result, not a measurement made for this note.

Available integration layers

OpenAI exposes Codex at three levels, depending on how much of the interface and runtime the host application needs to manage.

codex exec (headless / CI)

This fits CI/CD pipelines, cron jobs, and batch scripts. It runs a bounded, non-interactive workflow and can emit structured JSON without a user interface.

Codex SDK

The Node.js and Python client libraries let an application start, resume, and stream agent runs.

Codex app-server

This persistent local daemon uses a documented bidirectional protocol. It supports long-running conversations, streamed UI events, interruption, MCP tool discovery, and approval gates.

The Relay example

OpenAI open-sourced Relay, a sample logistics and supply-chain dashboard built on Codex app-server.

In the example, an operator starts from a delayed shipment record and asks to compare recovery options. The application exposes carrier rates and fleet availability through MCP. Codex can calculate alternatives, but a carrier rebooking requires approval before the write runs. The dashboard then updates the record.

Examples named by OpenAI

OpenAI's post names the following integrations and results. The wording here stays at the level the post supports rather than inferring specific interface surfaces.

  • GitHub and JetBrains: OpenAI identifies them as products integrating Codex.
  • Cisco: OpenAI says Cisco Cloud Control App Builder uses Codex for infrastructure generation.
  • Crete and Thrive Holdings: OpenAI reports more than 7,000 tax returns processed and a 33% reduction in preparation time.

Keep the domain interface

Relay is the part of the post I find useful. The operator stays in the logistics application, the selected record supplies context, and the existing approval control remains visible. Codex runs the loop; it does not require the host product to become a chat window.

Primary references