OpenAI opened the Codex engine: the best AI software may have no chat box

The bottom line

OpenAI has really open-sourced the core Harness that drives Codex Agent.

The CLI, official SDK and app-server can all be inspected, modified and embedded into your own products, and the openai/codex repository is licensed under the Apache-2.0 license. Developers can plug Codex's task loops, tool calls, approvals, and streaming progress into the business dashboard instead of creating a reskinned chat box.

But first step on the accelerator of "full open source": OpenAI open source is the Agent Harness and integration layer, not the Codex model weights, nor the complete Codex Cloud. IDE extensions and cloud-hosted services are still not open source, and model access is still another story.

A more accurate title would be:

OpenAI disassembles the Codex engine and gives it to you, but it does not send you the entire car and the oil field.

The all-purpose chat box has worn out its welcome

Over the past year, AI products have become more and more similar.

On the left is a column of history, in the middle is a chat box, and at the bottom is a line of "How can I help?"

Writing code is a chat box.

Checking data is a chat box.

Customer service, tax filing, operations, and supply chain were all shoehorned into the chat box.

The problem is, the people who actually do the work usually don’t rely on chat to understand the business.

The security staff looks at the alarm queue, the logistics staff looks at the delivery list, the customer service staff looks at the account history, and the financial staff looks at the reports and vouchers. The interface is more than just decoration, it is context itself, as well as boundaries of approval and responsibility.

The direction given by OpenAI this time is very straightforward: Stop driving everyone into Codex, let Codex enter the software they are already using.

What exactly is Harness?

A capable Agent is not equal to "model + a magical prompt".

It also requires an execution system:

• Understand and break down tasks;
• Preserve context across rounds;
• Read files and business data;
• Call terminal, MCP and other tools;
• Send progress to users in real time;
• Retry or change the path after failure;
• Request approval before hazardous operations;
• Finally write the results back to the original system.

The loop that wraps the model and keeps these actions going is Harness.

The model is like a brain, Harness is more like nerves, hands and feet, memory and brakes. The brain is very smart, but without these things, it still can only sit in the chat box and answer questions.

The strongest number: changing only the harness nearly tripled the score

OpenAI officials gave a very illustrative example.

In ARC-AGI-3 tests, GPT-5.6 Sol initially throws away previous private inferences at each execution step, and old context is truncated on scroll. The model looks like a man who briefly loses his memory with every step he takes.

OpenAI only adjusted two Harness settings:

• Retain the previous round of reasoning;
• Use contextual compression instead of simple truncation.

As a result, the score increased from 13.3% to 38.3%, and the output Token decreased by about six times.

This cannot be translated into "all tasks will be three times stronger", nor is it a new model benchmark. What it shows is: How the same model is managed may be more important than changing to a more expensive model.

Which three doors did OpenAI open this time?

First level: codex exec

Good for scripts, CI and one-off background tasks.

You can tell Codex to complete tasks within clear boundaries and return structured output. It does not require a complete product interface, but more like an Agent worker that can be plugged into an automated assembly line.

Second layer: Codex SDK

Suitable for applications launching, resuming, and streaming Codex tasks directly.

If developers need to use TypeScript or Python to manage threads, tasks, and events, the SDK is the more direct programming entry point.

The third layer: Codex app-server

This is the most noteworthy part.

Apps can connect to local Codex processes, maintain conversational state, receive events, interrupt tasks, expose their own tools, and handle human approval requests.

Business software continues to be responsible for interfaces, permissions, and data; Codex is responsible for Agent loops and controlled execution at the bottom level.

This means that you can have a logistics board that "works by itself" without having to put a chat window next to the board.

Official demo Relay: The chat box really disappears

OpenAI made a fictional logistics application Relay.

The user selects the delayed shipment and clicks "Compare Recovery Plans". The application automatically provides the manifest and operational data to Codex, and the Agent obtains the latest information by applying its own MCP tool.

If it is just analysis, it can directly give a plan; if you want to actually modify the booking record, approval must pop up. After the user approves, the tool performs writing and the Kanban board is refreshed synchronously.

The entire process does not require the user to write prompts from scratch.

Agent is not the protagonist of the product, but the execution capability hidden behind the button.

This is where this open source can really change software design.

Tax filing cases are more convincing than “making a demo”

OpenAI announces Thrive Holdings and Crete pilot to put Codex into tax preparation process.

The system processed approximately 7,000 tax returns in the pilot and reduced preparation time by approximately one-third. Modifications made by tax practitioners will also be recorded and converted into new assessment and engineering tasks, allowing the system to continue to improve.

This isn’t “Codex becoming a tax accountant yourself”. Professionals are still responsible for review, judgment, and rollout, and Codex is responsible for investigating failures, proposing modifications, running tests, and submitting candidates.

This boundary is very important:

Instead of letting AI file taxes independently, AI is connected to a tax system with data, rules, and expert approval.

Cisco also started to plug Codex into the cloud console

The official OpenAI article also lists Cisco's App Builder: it uses the Codex SDK to allow users to create applications through natural language in Cisco Cloud Control.

The point here is not that "Cisco made a chatbot", but that Codex was put into an existing cloud control product. Users still face the familiar business environment, and the Agent only undertakes generation and execution tasks behind the scenes.

If this model proliferates, many software may not have obvious “AI pages” in the future.

You'll just find that the original button can suddenly do things that used to require five people to communicate back and forth.

The reversal: this is neither model open-sourcing nor a free lunch

OpenAI official documentation clearly distinguishes the scope of open source:

• Codex CLI: open source;
• Codex SDK: open source;
• Codex app-server: open source;
• Codex Security CLI/SDK: has open source components;
• IDE extension: not open source;
• Codex Cloud: not open source;
• Model access and hosting services: separate from open source Harness.

So you can inspect and modify the execution layer of "how the model works" but still need to address model invocation, deployment costs, certification, and cloud services.

Apache-2.0 allows modification and commercial use. It does not mean that calling OpenAI models is free of charge, nor does it mean that all Codex product codes have been made public.

DeepSeek faces the most awkward—and potentially happiest—reaction

There is also a very interesting subplot to this matter.

DeepSeek has officially provided tutorials for connecting your own models to Codex. Now that Codex Harness is more clearly promoted as an open platform, in theory it is easier for developers to combine OpenAI's Agent shell, their own business interfaces and other models.

So a very abstract product may appear:

The interface is yours, Harness comes from OpenAI, but the model is DeepSeek.

This isn’t necessarily a bad thing for OpenAI. As long as Codex Harness becomes the default Agent running layer for developers, even if some people change their models, OpenAI will still be able to define the interface and working methods of the entire ecosystem.

The real competition begins to shift from "whose model scores the highest":

Who can become the Agent operating system that all models are willing to live in.

My judgment: This is not zero-friction; it means building one less wheel from scratch

The most easily exaggerated word in the original manuscript is "Agent zero-threshold era."

Embedding agents into enterprise software still requires permission design, MCP tools, logging, evaluation, error recovery, cost control, and manual approval. Harness open source won't make these problems go away automatically.

But it does eliminate a huge chunk of duplication of effort: each team no longer has to implement session state, event streams, tool calls, sandboxing, and approval protocols from scratch.

In the past, when developers made AI products, they often made a chat box first; next, developers might first ask:

Which interface does the user originally work on? Which action should the Agent hide behind?

This is a healthier product direction than "putting on another shell".

What to watch

• Will the community build new front-end and vertical industry products around app-server?
• What is the compatibility and effect of non-OpenAI models after they are connected to Codex Harness?
• Will the functional gap between Apache-2.0 open source components and closed source Codex Cloud widen?
• Whether the enterprise is willing to hand over high-privilege tools to local Codex processes.
• The Harness ecosystem will ultimately be dominated by OpenAI, Claude Code, DeepSeek or independent open source projects.

The last sentence:

In the past, all software wanted to add a chat box; in the future, the best AI software may not see a chat box at all.

Codex Harness FAQ

Has OpenAI open-sourced the Codex model?

No. Open source is Codex Harness and related integration components, including CLI, SDK and app-server. Model weights, model access, and hosting services are not open source as a result.

Can Codex Harness be used commercially?

The openai/codex repository is licensed under the Apache-2.0 license, which generally allows modification, distribution, and commercial use, but is still subject to the terms of the license, third-party components, and usage rules of the called model services.

What is the use of Codex app-server?

It lets applications connect to local Codex processes, manage threads and tasks, stream events, invoke application tools, and handle human approval before dangerous operations.

Will it be free to use Codex after it is open source?

no. The open source code is available for free, but there may still be fees for model APIs, computing resources, hosting services, and actual deployment.