Codex is no longer only a terminal tool that writes code. Its 2026 product line includes a desktop app, CLI, IDE extension, cloud tasks, Skills, Plugins, and an Agents API for developers. The names are easier to understand when divided into three layers: the Codex interfaces you operate, reusable extensions for team workflows, and an API for putting the same agent harness inside another product.
Start with the four user interfaces
Codex app
The desktop app is a control center for multiple agents. Tasks live in separate threads, and Git worktrees isolate changes so two agents do not have to operate in the same working directory. OpenAI added a Windows version in March 2026; current operating-system support and downloads should still be checked on the official product page.
The app is useful when one agent fixes a bug while another writes tests and a third checks documentation. More agents are not automatically better. Each thread still needs a defined scope and an acceptance test.
Codex CLI
The CLI works directly in a repository from the terminal. It can inspect files, make edits, and run commands while using project instructions, Skills, MCP servers, and sandbox controls. Its permission configuration determines which paths it can access, whether it can use the network, and which commands require approval.
IDE extension
The IDE extension is designed for working beside the code, selecting relevant files, reviewing diffs, and continuing the current development context. It is a different interface, not evidence that the underlying model is automatically stronger than it is in the CLI.
Codex cloud
Codex cloud runs a task in an isolated cloud environment. A user can connect GitHub or GitLab, configure dependencies and environment settings, let the task continue in the background, and then review its summary and diff or open a pull request. It is useful for longer jobs, parallel attempts, and work that should not occupy a local machine.
Skills and Plugins solve different problems
A Skill packages instructions, resources, and scripts for a repeatable job. It can describe a deployment procedure, documentation standard, data-analysis workflow, or UI validation process so Codex does not have to rediscover the team's method every time.
A Plugin can combine apps, skills, instructions, and workflows. In June 2026, OpenAI announced role-specific plugins for data analytics, creative production, sales, product design, public-equity investing, and investment banking. The intent is to connect Codex to non-engineering tools and data, not merely to add another prompt template.
Neither mechanism makes bad inputs reliable. Incorrect permissions, source data, instructions, or scripts can cause an agent to execute the wrong process efficiently. Review the data source and allowed actions before sharing a Skill or Plugin.
What are Sites and annotations?
Sites let Codex turn analysis or plans into an interactive website or lightweight tool that can be shared within a workspace. When OpenAI announced the feature, it was a preview for Business and Enterprise customers, so individual availability should not be assumed.
Annotations let a user point to the exact part of generated code, Markdown, a document, spreadsheet, slide, or site that needs refinement. They are intended for targeted iteration after a first draft instead of restating the entire request.
The Agents API is not the Codex app
OpenAI launched the Agents API in public beta on September 10, 2026. Developers can create a cloud agent session by choosing a model, tools, environment, and subagent settings. OpenAI manages the Codex-derived harness for context, recovery, long-running sessions, and intermediate results.
The practical distinction is straightforward:
| Need | Product to evaluate |
|---|---|
| Ask an agent to modify your repository | Codex app, CLI, IDE, or cloud |
| Standardize a repeatable team workflow | Skills or Plugins |
| Add an agent to your own SaaS or backend | Agents API |
Public beta products can change quickly. OpenAI currently states that the Agents API has no separate platform fee and that users pay for the tokens and tools they consume. Check the current beta documentation and pricing before a production deployment.
Codex is expanding beyond engineering
OpenAI reported in June 2026 that Codex had more than five million weekly users, with non-developers accounting for about 20 percent and growing faster than developers. Reported knowledge-work use cases include reports, spreadsheets, presentations, research, data analysis, and workflow automation.
That does not remove the need for data boundaries. A non-engineering workflow often needs stricter scoping because it touches customer, financial, or company information. A sales-report task may need access to one approved dataset, not write access to an entire drive or production database.
What should a new user try first?
- Use the app or CLI for one small task with a clear acceptance test.
- Add project instructions and check whether the second attempt is more consistent.
- Turn a workflow into a Skill only after it has been repeated several times.
- Use cloud tasks when work must run in the background or when several approaches need comparison.
- Use the Agents API only when the agent must become part of your own product.
This sequence makes time savings and costs visible before permissions and automation expand.
Related comparison: Codex vs Claude Code vs Antigravity vs Copilot CLI
Official sources
- OpenAI: Introducing the Codex app
- OpenAI: Codex cloud
- OpenAI: Codex for every role, tool, and workflow
- OpenAI: Introducing the Agents API
- OpenAI: How agents are transforming work
繁體中文： Codex 2026 新功能整理