Claude vs Codex: which one for your team?

Claude is the AI assistant of Anthropic, and developers use it as a coding agent. Codex is the coding agent of OpenAI. Both read a repository, change files, run commands and propose changes for review. Which one fits a team depends on the team’s code, its data rules and its review process. This article names no winner. Prices, limits and model versions of both products change within weeks, and a verdict written today is soon out of date.

What Claude and Codex have in common

A coding agent is a model plus its agent harness. The harness is the running setup around the model, with its prompt, tools, hooks and loop (LangChain, March 10, 2026). Claude and Codex both come with a harness.

Both read written instructions from a file in the repository. Codex reads AGENTS.md, and Claude reads CLAUDE.md. One file can be the source that the other imports. A rule is then written once and holds for both tools.

Both follow a written rule most of the time. For code that goes to production, most of the time is too little.

Where the two differ for a team

A hook binds only the harness that loads it. A team that works with both tools writes its hard checks twice, once per tool.

Three questions before a team picks one

The first question is the code. The agent has to fit into the repository, the build and the review process the team has today. The team writes down which directories the agent may change and which commands it may run.

The second question is the data rules. The company decides which data an agent sees, and its data protection officer decides which data may go to a cloud model. The company’s plan with the vendor and its settings belong to this question.

The third question is the review. Generated code reaches the main branch through a merge request, as a colleague’s code does. The team names the developer who reviews it and the checks that must pass first: tests, linters and a build.

A team that can answer the three questions for one tool can answer them for the other. The answers move with the team when it changes the tool.

The rules and the review decide the result

Limits at the resource hold in any harness, because the database and the repository enforce them. The agent gets a database role that only reads and tokens scoped to development. Its credentials cannot push to the production branch.

A rule whose break would do large damage becomes a hook that blocks the call before it runs. The hook gets tests for the calls it must allow, because developers switch off a check that blocks valid work. A write to live data waits for a person’s confirmation, given outside the chat.

How I work with both tools

I am an IT expert for data platforms in Frankfurt. For a client I break up a monolithic platform with Claude and Codex. The project has more than 300 RESTful API endpoints and more than 8,000 commits. The two agents work in one repository under the same rules. A code agent works in its own git worktree, commits under its own name and opens a merge request. The data model and the review of the results are my part of the work.

As a Claude consultant in Germany I set the tool up for a team’s developers and write the rules for the code it generates. A team that already works with Codex keeps it. My Claude training in Frankfurt has a separate session for developers on Claude as coding agent. In the ChatGPT and Codex training in Frankfurt developers learn Codex on their own code base. In the workshops the participants work in the tools on their company’s material. I take on this work as a freelancer or as a permanent employee.