Big move! OpenAI fully open-sources Codex Harness

OpenAI has open-sourced the three core components of Codex Harness (CLI tools, official SDK, and app-server) under the Apache 2.0 license. Harness serves as the underlying execution framework driving AI agents, allowing seamless integration into developers' own products and completely moving away from the "generic chat box" model. It also returns interface, data, and security approval rights to developers, effectively bridging the final mile for enterprise-level AI deployment.

Just yesterday, OpenAI completed another upgrade.

This wasn't a completely new model release, but its impact on developers was even greater than the iterations of the model itself.

OpenAI has officially announced that it will fully open-source the renowned Codex underlying core framework (Harness) as part of its platform!

In other words, OpenAI is giving away the engine that powers its top-tier AI agents for free.

Developers no longer need to cram business processes into a generic chat box!

You can directly integrate the most powerful AI agents into your own products, engineering tools, operational dashboards, and even the most mundane financial reporting software.

OpenAI President Greg Brockman excitedly retweeted on Codex, saying: "Codex can power so much more than just programming tools!"

You can build your own product directly on Codex Harness.

From now on, you're in charge of your own space. The interface, data, and permission approvals are all yours, while AI is only responsible for doing the "work" at the bottom level.

How will OpenAI's latest "open source mega move" completely disrupt the future software development paradigm?

The world has long suffered from the "universal chat box".

Think back to how we use AI now?

Whether it's writing code, doing data analysis, or handling customer complaints, the standard procedure seems to be: open a webpage/sidebar, face a monotonous chat input box, type in a background description, and then expect AI to provide a perfect answer.

If the answer is wrong, continue to "pull and tug" in the chat box.

Most people are familiar with Codex through its apps, command-line interface (CLI), or IDE extensions.

But OpenAI believes that this single interactive experience is far too narrow!

OpenAI succinctly pointed out in its official blog:

Instead of requiring each team to forcibly migrate their familiar workflows to a generic code assistant, it's better to integrate the Agent directly into software designed around real-world tasks.

After all, security analysts look at the alert queue and the status of affected services; customer service engineers look at account history and product logs; and product managers look at the requirements dashboard.

For these real workers, what matters isn't the "chat box," but the business dashboard in front of them. The interface itself is meaningful; it is the most important "context."

Now, with the open-source release of Codex Harness, this revolution of "anti-shelling" has officially begun.

Unveiling Harness: The "Most Powerful Exoskeleton" for AI Agents

So, what exactly is Harness , the main subject of this open-source project?

Many people mistakenly believe that a powerful AI agent equals a good model plus a good prompt.

Absolutely wrong!

A truly functional intelligent agent requires an extremely complex underlying execution system.

It needs to understand the task, retain memories during lengthy conversations, review relevant information, skillfully utilize various tools, present progress to the outside world, handle crashes and failures, stop at critical moments to request human approval, and finally return useful results.

This "execution system" and "agent loop" that handles all the dirty work is Harness.

With this open-source release, OpenAI has proven one thing to the outside world: a good Harness design can even transform a model.

OpenAI provides very compelling data to demonstrate how critical harness design is.

In the extremely challenging ARC-AGI-3 benchmark test, they made only two key adjustments to Harness—preserving inference and context compression.

As a result, the score of the GPT-5.6 Sol model instantly soared from 13.3% to 38.3%! Even more astonishingly, the number of output tokens decreased by a factor of six!

In other words, with Harness's support, AI not only becomes three times smarter, but also saves you a huge amount of API call costs.

As it turns out, while model capabilities are important, how to manage the model (i.e., Harness's design) is the key to determining the agent's final performance.

Three open-source components, zero-barrier integration

To enable developers to truly "use it out of the box", OpenAI released three complete open-source components under the Apache-2.0 license.

You can unlock this powerful arsenal simply by visiting the openai/codex repository on GitHub.

https://github.com/openai/codex

First is the CLI tool codex exec, which can run automated pipelines.

If you only need the AI to run a script, execute a CI (Continuous Integration) task, or run a one-off background task, use the CLI directly.

It can run a bounded agent workflow and return structured output. Simple, straightforward, and efficient.

The second is the official Codex SDK, which is the ultimate joystick for programmers.

If you are writing applications that need to start, resume, or stream Codex tasks using code, the official SDK (which supports TypeScript/Python) provides a direct programmable interface.

You can precisely control the lifecycle of threads and tasks using code.

The third, and most dazzling star of this open-source release, is Codex app-server. It is the core engine that is fully integrated into the product.

When an agent needs to be part of your product, app-server is the best choice. It allows your application to connect to the local Codex process by recording a detailed client protocol (JSON-RPC).

You can use it:

Maintain a sustained conversation;

Streaming events (see what the AI is doing in real time);

Interrupting the AI's work midway;

Expose the tools of your own application to AI for use;

Processing human approval requests (Human-in-the-loop).

As renowned developer @ClusterProtocol lamented:

By separating the front-end approval interface from the underlying execution loop, embedding the Agent into the business dashboard has become simpler than ever before!

It's much more than just writing code! It's a "nuclear weapon" for efficiency improvement across all industries.

If you think Codex is only for writing code, then you are very wrong.

As Brockman said, Codex can drive much more than just programming tools.

Let's take a look at what miracles those pioneers have created with it:

Case Study 1: A "God-Level Assistance" for Tax Filing

Can you imagine AI filing taxes? The biggest fear in accounting work is making mistakes; the margin for error is extremely low.

Partners Thrive Holdings and Crete leveraged this open-source Harness suite, embedding it into a professional tax preparation workflow.

The system can not only handle complex tax logic, but also perfectly integrate feedback from tax advisors.

The results were astounding: the pilot project successfully processed as many as 7,000 tax returns and reduced tax preparation time by about one-third!

This demonstrates the immense power of embedding AI into specialized vertical tools.

Case Study 2: Cisco's Cloud Management System

Tech giant Cisco uses the Codex SDK to build App Builder within its cloud control platform.

Now, Cisco customers can directly use natural language to create custom applications, while the complex underlying logic is handled silently behind the scenes by Codex Harness.

Official demo template: Logistics Kanban "Relay"

To help everyone understand more intuitively, OpenAI officially showcased a virtual logistics operations dashboard application called Relay .

There are no distracting "ChatGPT chat boxes" on this board.

1. Dramatic change in interaction : Users no longer need to write a Prompt from scratch. They simply select a delayed invoice and click "Compare Recovery Options".

2. Seamless context integration : The application will automatically feed the AI with the current page's order details and logistics data as context.

3. MCP Integration : Codex will automatically call the application's own MCP tool to obtain the latest real-time operational data.

4. Human oversight : After the agent analyzes and determines the optimal rebooking plan, an approval dialog box must pop up. The AI will only execute the write operation after the salesperson clicks "agree".

5. Status synchronization : After the operation is completed, the business dashboard will be automatically refreshed.

In this process, Harness handles the agent loop, memory maintenance, tool interaction, and real-time feedback; while your product retains control over its dashboard, data recording, and overall management. This is true "human-machine collaboration."

Reshaping boundaries, returning power and soul to application developers

Why do we say that this open source has changed the paradigm of software building?

Over the past year, many developers have fallen into a vicious cycle of "shelling" products. Everyone is making ChatBots, products are becoming increasingly similar, and competitive advantages are becoming increasingly shallow.

As an open platform, Codex gives developers "supreme control" in three dimensions:

Codex gives users absolute control over the interface; stop forcing users to adapt to the chat box!

Let your users continue using the dashboards, editors, maps, and lists they are familiar with. AI should be an invisible helper, not a screen-dominating protagonist.

Absolute control over context and tools gives you peace of mind to directly expose your company's most critical systems, confidential documents, and internal APIs to the Agent through the MCP service.

AI is no longer an outsider who "doesn't know the company's rules and regulations," but an expert who is proficient in your business systems.

Finally, absolute operational boundaries and security control are crucial for enterprise applications. The host application can determine where the agent runs and what files it can access.

Most importantly, you can set which dangerous actions (such as modifying a database or sending emails to customers) must be authorized by a human click.

This model, where "the front-end business rules belong to you, and the underlying Agent is cyclically returned to OpenAI," completely bridges the last mile in the implementation of AI in enterprise-level scenarios.

Codex harness is licensed under the Apache 2.0 license, meaning anyone can modify and use it commercially.

The real test is: how many developers will abandon the comfortable chat box and disassemble the intelligent agent into parts to fit into their products?

This product philosophy is deeply embodied in the Codex harness.

What we need to do is not to use a universal chat box to eliminate all professional interfaces; rather, we need to equip these interfaces with a smart brain.

When generic chat boxes disappear, native AI applications will flourish.

The era of zero barriers to entry for I Agents has arrived.