codex

this is codex

a screenshot of me working on some personal projects and some content for a brand deal.

Codex is OpenAI’s agent for software development and technical work. it reads project context, works through tools, changes files, and checks the result. you can use it through the terminal, an editor, the desktop app, or a cloud environment. the surface changes how you steer and inspect the work.

on the spectrum of how personified these products feel to me, Codex sits somewhere between Claude Code and Grok Bot. the main tasks get descriptive titles, but then i’ve seen a subagent show up named Casper or Socrates. some fucking random names. that small detail gives the agents a bit of personality alongside the work they’re doing.

where can i use Codex?

terminal and editor

the CLI puts the conversation beside commands, output, and the current directory. an editor adds selected code, diagnostics, and inline review. both are useful for a change whose behavior and verification are already clear.

for example, “find why this test fails and fix the implementation” gives the agent a specific path through the repository. inspect the changed code and rerun the command after the fix. the interface helps you see the work; the check explains what the change actually achieved.

desktop and parallel tasks

i use one authoritative Codex thread to keep track of multiple projects, tasks, and contexts running in parallel. it turns what i want into clearer instructions and routes them to the right tasks. i explain that setup in the delegation section.

the desktop app brings conversations, terminals, previews, and review together. worktrees give independent changes separate checkouts. each still needs an owner, a clear outcome, and any local setup needed to run its checks.

this February 2026 walkthrough shows the Codex app’s approach to supervising several tasks.

watch the Codex walkthrough ↗

OpenAI Developers (@OpenAIDevs), february 2026.

cloud work and mobile access

cloud tasks use hosted environments with repository setup and dependencies. Remote lets a phone or another supported device steer work on a connected host. the host retains its files, credentials, tools, and permissions. an SSH project uses the remote machine’s environment.

check where a task runs before choosing how to control it. a phone can be convenient for answering a narrow question; a wide diff may be easier to review on a larger display. a hosted task needs its own setup even when the same repository already works on your laptop.

how does Codex run a task?

the app server exposes the agent loop to clients. a thread holds a conversation, a turn represents a request and the work it causes, and items describe messages, commands, tool calls, or file changes. the client receives events as that work progresses and responds to approval requests when required.

this is useful when building an interface or coordinator around Codex. reading history shows what happened; current status and incoming events show how work is progressing. the client can present those details and route the next instruction to the appropriate thread. use the schema generated by the installed version when an integration depends on particular fields.

instructions, tools, and permissions

repository instructions provide conventions and verification commands. skills provide procedures. MCP connections provide operations in other systems. hooks run handlers at defined events. subagents handle delegated parts of a task. permissions constrain the actions available to the running session.

for a concrete example, an instruction file can point to the project’s check command, a review skill can explain what to inspect, and a subagent can investigate a particular failure. their roles are explained in extensions and configuration.

can i change Codex itself?

OpenAI publishes the CLI, SDK, and app server as open source. that makes source changes possible alongside ordinary configuration and extensions. changing source involves building or running that version; editing a config file adjusts the behavior exposed by the installed version. the cloud service and IDE extension have separate availability in OpenAI’s component list.

how do i check the result?

a useful task ends in something you can evaluate: a reproduced bug, an explained code path, a patch with checks, or a working interface you can use. name that result in the prompt. include constraints that change the work, such as preserving an existing API or keeping another task’s files intact.

then inspect the result at the level the task requires. passing unit tests support the behavior they exercise. a visual change needs a browser check. a pushed branch still needs review and integration before it becomes the version users see.

the configuration guide explains what belongs in project instructions and what memory carries forward. the coordination example shows how one task can route work to others and check what comes back.