What is Codex? Distinguishing the coding agent, clients and underlying models
How OpenAI Codex reads projects, edits code and uses tools, with distinctions between GPT models, completion, local execution, permissions and verification.
Meet this tool
Codex is OpenAI’s coding agent for writing, reviewing and debugging software with project files. Its clients and harness organize models and tools: for example, reading files related to an error, editing the implementation and running existing checks. Codex as a product does not mean one fixed GPT model. Models, execution locations and permissions depend on the interface and configuration.
What the software actually does when repairing a project
For an unresponsive login button, Codex can inspect related components and calls and propose or apply edits. Tools return files and command results to the model, informing its next tool call. It does more than continue code at the cursor, but task records are needed to establish whether checks actually ran.
GPT and other models interpret information and produce decisions; Codex organizes execution around engineering work. Changing the model affects behavior, while software still manages repository access, command tools and task state. Attributing every product result to the model overlooks environment and tool configuration.
One product can provide several working interfaces
Codex has command-line and editor interfaces and can participate in remote tasks or programmatic workflows. The CLI works with a selected project and available machine tools. Other execution arrangements can differ in file location and dependencies. A locally started client does not mean all computation occurs locally.
Projects can use AGENTS.md for build commands, coding conventions and files to preserve. Codex reads applicable instructions before work, with directory hierarchy affecting scope. These are project instructions for the model, not filesystem access controls; the execution environment still determines accessible directories.
Model suggestions and environment execution are separate
The execution environment and permission policy constrain writable files, network access and command approval. A proposed command is not necessarily approved or successfully executed. Repositories may contain confidential material; provide only necessary access instead of assuming a coding product can use the entire machine.
External tools add possible actions, but account authorization and resource permissions still apply. Sandboxes constrain operations; they do not guarantee correct code logic. Plausible terminal explanations do not replace actual process results and file differences.
Review code differences and actual execution results
Completion fits a known editing location. Coding agents suit tasks needing cross-file investigation, commands and feedback. Choose interfaces for existing workflows, available tools and review requirements rather than product names or model lists alone.
Inspect actual changes, run risk-appropriate checks and record untested environments. Official documentation establishes mechanisms, not success on every task. This page provides no independent success-rate or speed ranking. Check current configuration for available models and account eligibility.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.