A client of mine runs a $7M distribution business on eleven spreadsheets. Last month he forwarded me an article about OpenAI Codex with one line attached: "Does this mean I can finally get my order tracker built?" It is a fair question, and the honest answer takes more than a headline. Codex really does write software. It fixes bugs, builds features, and works unattended for hours. Whether it can build an app for a business owner who has never written a line of code is a different question, and the answer lives in the details of what Codex was designed to plug into.
I am going to give Codex its full due here, because it deserves it. Then I want to show you the part the headlines skip: the workflow it assumes you already have, and what the same class of AI actually looks like when it is packaged for someone running estimates, orders, or client onboarding out of Excel.
What OpenAI Codex Actually Is
Codex is OpenAI's agentic coding system, and as of this writing it comes bundled with ChatGPT plans. You can reach it three ways: a command-line tool developers run in a terminal, an extension that lives inside code editors like VS Code, and a cloud agent inside ChatGPT that takes on longer tasks in the background. Give it work and it clones a copy of your code repository into an isolated environment, reads the codebase, edits files, runs the test suite, and comes back with a set of proposed changes plus logs you can trace step by step.
The GitHub integration is genuinely impressive. Connect a repository and the cloud agent can pick up a task, do the work, and open a pull request, which is the formal package developers use to propose changes for a teammate to review. Engineers can even tag the agent on an issue and it starts working without anyone leaving GitHub. Teams report handing it routine refactors and migrations and getting back days of work overnight.
None of that is hype. For a software engineering team, this is the biggest productivity shift in a generation. My client's instinct was correct: the technology to build his order tracker absolutely exists now, and it is not even expensive.
Look at the Units Codex Thinks In
Read the last section again and notice the vocabulary. Repository. Branch. Pull request. Code review. Test suite. Terminal. Every one of those words names a piece of an engineering workflow, and Codex is brilliant precisely because it slots into that workflow so cleanly. It behaves like a tireless new engineer who joined your existing team and follows your existing process.
Now ask the question that matters for your business: which existing team? Which process?
When my distribution client asked whether Codex could build his order tracker, here is what the unassisted path would actually require of him. Create a code repository, an account and a structure he has never touched. Describe the app well enough, in software terms, for the agent to start. Then receive the output, which arrives as a pull request full of code, and review it, because the agent's whole design assumes a qualified human checks the work before it ships. Then stand up hosting, a database, backups, user accounts, and security. Then keep all of it patched and running, forever, including the morning it breaks while forty orders are in flight.
He runs a warehouse. He is exceptional at it. Every hour he spends learning what a branch is comes out of the business that pays for everything else. The tool is not the barrier. The workflow around the tool is, and that workflow exists for good reason: professional software teams need coordination, review, and version control. A twelve-person distributor does not need any of that. He needs his spreadsheet to become software, and he needs someone accountable for keeping that software alive.
What a Spreadsheet-Run Business Actually Needs
I have sat with dozens of operators at this exact decision point, and the requirement list is remarkably consistent. The orders tab needs to become a real database so two people can update it at once without silently overwriting each other. The lookup formulas need to become screens a new hire can use on day one. The rules everyone honors on good days, credit check before the order ships, deposit before the crew is scheduled, need to become gates the system enforces on bad days too. I walked through why the database layer changes everything in using Excel as a database and the transition to custom software.
Notice that nothing on that list is a pull request. The operator's requirement is an outcome: working software, hosted somewhere reliable, that keeps matching the business as the business changes. The engineering workflow that Codex serves so well is a means to that outcome for companies that employ engineers. For everyone else it is a wall between them and the outcome.
Codex Alone vs. the LlamaPress Path
Here is the side-by-side I drew for my client. This is a comparison of packaging, and I want to stress that, because under the hood both paths use the same class of technology: a frontier AI coding agent doing real software work.
| OpenAI Codex, Used Directly | LlamaPress + Leonardo | |
|---|---|---|
| Designed for | Engineering teams with an existing codebase and review process | Operators whose process lives in Excel or Google Sheets today |
| Starting point | A code repository you set up and connect | The spreadsheet you already run the business on |
| How you request work | Task prompts framed in software terms, via terminal, editor, or ChatGPT | Plain English, the way you would brief a contractor |
| What you review | Pull requests: diffs, logs, and test results a developer signs off on | The working app itself; you click through screens, not code |
| Hosting & upkeep | Yours to arrange: servers, database, backups, security, patching | Managed hosting included; humans on call when something breaks |
| Skill assumed | Someone on staff who can read code and own an engineering workflow | You know your own process; that is the whole requirement |
When Codex Is the Right Tool
Plenty of readers should close this tab and go set up Codex, and I would rather say so plainly than pretend otherwise. Use Codex directly if you employ developers, even one or two, because it will multiply them. Use it if you already have a codebase, since an agent that lives in your repository and opens pull requests against it is exactly what an existing product needs. And use it if you have, or want to build, a real engineering workflow with version control and code review, because that discipline is what makes unattended AI coding safe at scale. A software company of any size that skips Codex or a tool like it in 2026 is leaving money on the table.
The dividing line is honest and simple. Codex assumes the engineering context already exists. If it does, the tool is superb. If your company's entire technical footprint is a shared drive and some very hard-working spreadsheets, then adopting Codex means first building an engineering function to wrap around it, and that was never your job.
The Same Class of Agent, Packaged for Your Business
This is exactly the gap LlamaPress was built to close. Our AI coding agent is named Leonardo, and it does the same kind of work the headlines describe: it writes real code, builds real database-backed applications, and iterates for as long as the job takes. The difference is everything wrapped around it. You upload the Excel file that runs your process, and Leonardo reads the tabs, columns, and formulas as the specification. Your customer sheet becomes a customers table in a genuine relational database. Your quoting logic becomes tested code. Your follow-up column becomes a queue that will not let an order fall through the cracks.
The app arrives on managed hosting, already running, with backups and security handled. When the business changes, and it will, you describe the change in plain English: "add a rush flag to orders and put rush jobs at the top of the picker's screen." Leonardo makes the change. Humans at LlamaPress stay on call for the moments that need judgment. You review the result the way you review any contractor's work, by using it, and your job title never has to include code reviewer.
For the full picture of how a spreadsheet becomes an application, step by step, the Excel to app master guide walks the entire journey. If you are weighing this path against the low-code platforms, I have covered that ground too, in the Power Apps alternative for businesses without an IT department and in why an AI app builder is the only way to replace your spreadsheet.
Your spreadsheet is the spec. Leonardo does the rest.
Upload the Excel file that runs your orders, estimates, or onboarding. Get back a database-backed app on managed hosting, with changes made in plain English and humans on call.
Frequently Asked Questions
Can I use OpenAI Codex to build an app without knowing how to code?
Realistically, no, and the limitation is the workflow rather than the AI. Codex delivers its work as code changes inside a repository, packaged as pull requests meant for a developer to review, and it leaves hosting, databases, and maintenance to you. A non-developer can get fragments working, but shipping and running a real multi-user business app that way requires engineering skills at several steps.
What is OpenAI Codex for, then?
Codex is built to make software engineering teams dramatically faster. As of this writing it works through a command-line tool, editor extensions, and a cloud agent in ChatGPT, and it shines at tasks like building features in an existing codebase, fixing bugs, refactoring, and opening pull requests on GitHub for engineers to review.
Is there a Codex alternative for non-developers?
The practical alternative is the same class of AI coding agent delivered as a service instead of a tool. With LlamaPress, the agent Leonardo builds your application from your uploaded spreadsheet, the app runs on managed hosting, you request changes in plain English, and humans stay on call. You never touch a repository or review code.
Can AI really replace my order-tracking spreadsheet?
Yes, and this is precisely the job AI coding agents are suited for. Your spreadsheet already encodes the process: the columns are your data model and the formulas are your business rules. An agent reads that as a specification and produces a database-backed app with real user accounts, enforced workflow rules, and a single source of truth your whole team shares.
Do I need GitHub or ChatGPT to use LlamaPress?
No. The LlamaPress flow starts with the spreadsheet you already have. You upload it, Leonardo builds the application, and everything after that happens in plain English inside your app. There is no repository to create, no plan to subscribe to on another platform, and no pull request will ever be waiting for your sign-off.