AOSplaneCompareClaude Code vs Codex
Claude Code vs Codex for autonomous pull requests
Claude Code and Codex both run on AOSplane as cloud agents, on accounts you already pay for, with no markup. You choose the harness and the model. AOSplane does not pick one for you: matching a task to an agent automatically is still being built.
- Hard spend cap per task
- Linear and Jira
- Claude Code and Codex
6 sections · 3 min read
On this page6 sections
The short version
Claude Code is Anthropic’s coding agent. On AOSplane it reads the repository, including a CLAUDE.md, and is often used to build and to review. Codex is OpenAI’s coding agent. The Codex CLI is the one AOSplane runs. It reads an AGENTS.md, and it fits tightly scoped work with a short loop, including writing tests and fixing a small diff.
The controls around them are the same: one isolated sandbox per run, a budget reserved before the run starts, an approval when the agent asks to do something sensitive, and a record of the run. A draft pull request is how the change comes back. Agents cannot push, and a person merges.
Claude Code
Claude Code plans a change, edits across files, runs the tests and opens a pull request. Hooks, subagents and MCP servers extend it. On a laptop it is one session at a terminal. On AOSplane it runs in the cloud, so more than one can run at once, and a teammate can start it from a Linear issue, a Jira ticket or the API.
You connect an Anthropic API key, Amazon Bedrock, Claude Platform on AWS, another compatible provider, or a Claude subscription where sign-in is available. Anthropic’s charges and limits stay on that account. The sandbox clones the repository, so the CLAUDE.md there still applies. The full guide is Claude Code as a cloud agent.
Codex
Codex is a family of coding agents. The CLI is the open-source terminal version. OpenAI also runs Codex inside ChatGPT, in OpenAI’s own sandboxes, and as an IDE extension. If you only ever run Codex, OpenAI’s hosted option is the shortest path. Teams use AOSplane when Codex is one of several agents, or when the budget, the gate and the record matter as much as the run.
On AOSplane you connect an OpenAI API key, Azure OpenAI, Amazon Bedrock with a bearer token, another compatible provider, or a ChatGPT subscription. The sandbox works from your repository, so AGENTS.md applies. The full guide is Codex as a cloud agent.
Side by side
| Claude Code | Codex | |
|---|---|---|
| Whose agent | Anthropic | OpenAI. AOSplane runs the Codex CLI. |
| Instructions in the repo | CLAUDE.md | AGENTS.md |
| Where it fits | Build, and review, including changes that need more of the repository. | Short loops: write the tests, fix a flaky check, a small diff. |
| Account on AOSplane | Your Anthropic account, Bedrock, Claude Platform on AWS, or a Claude subscription. | Your OpenAI account, Azure OpenAI, Bedrock, or a ChatGPT subscription. |
| Sandbox, budget, gate, record | The same as Codex | The same as Claude Code |
| Who chooses it | You, per agent | You, per agent |
Using both on one task
A task can use more than one agent, one role each. The example on the home page is Claude Code to build, Codex to test, and Claude Code to review. That split is not a rule. Each role is its own agent, with its own harness, model, instructions and budget, and each stage hands the next one context that has been checked.
With Plan first, the product manager writes the spec before either agent writes code, and you approve it. QA then runs the repository’s checks and one command per acceptance criterion. The draft pull request opens after they pass. Kiro, Cursor and Tower can sit in the same fleet, in preview, and might not be enabled for every organization.
What this page does not say
There is no score here for which agent writes better code, and no claim that one is faster. Those would be measurements this site does not have. The difference above is the one the product guides already draw: which provider, which instructions file, and which kind of task each is aimed at. You can change the assignment per agent.
Read the cloud agents guide for the loop around either harness, or how a ticket becomes the pull request.