OpenAI Codex

Codex is OpenAI’s agentic coding tool, and its defining trait is reach. The same agent runs in the terminal, inside an editor, in a desktop app, in the browser and on cloud machines, so a task started in one place can be picked up in another. It is the most widely distributed of the coding agents, and the cheapest to try.

What it is

Codex takes a goal rather than a prompt-and-response exchange. It gathers the context it needs, acts on the codebase, and produces something you can review: a set of edits, a review of a pull request, a fix with the tests already run. The command-line version inspects, edits and runs code against a local repository, so the work happens where the code already lives. Alongside the coding surfaces sit the wider assistant features, including code review of pull requests, web search, scheduled tasks and connections to GitHub, Slack and Linear.

How it works

Underneath is the GPT-5.6 family, split into tiers with different trade-offs: Sol for the hardest reasoning, Terra for balanced everyday work, and Luna for speed and lower cost. The tier is chosen per session with a /model command, so an engineer can spend heavily on a difficult refactor and cheaply on routine edits.

The safety design is worth understanding, because it is the part users interact with most. Codex runs inside a sandbox with a defined set of writable directories, and a /permissions setting controls when it may edit files or run commands without asking first. You can inspect the active sandbox and the writable roots before letting a run continue. Like the rest of the field, it connects to outside systems through Model Context Protocol servers.

Where it fits

Codex is the generalist of the coding agents. Anthropic’s Claude Code is usually described as the deeper programmable harness, and Cursor as the stronger in-editor experience, while Codex spans the most surfaces at the lowest entry price. In practice many teams run more than one, using whichever suits the task rather than standardising on a single tool.

Where it stands

The distance between the leading agents narrowed sharply through 2026, and Codex closed most of the gap it started with. Recent additions have pushed towards continuity and parallelism: conversations now sync across devices, and the tooling has moved towards running several agent sessions at once rather than one at a time. The unresolved question is the same everywhere in this field, which is how much a team lets an agent do unsupervised, and the permissions dial is OpenAI’s answer to it.