今日热榜中文

Codex vs Claude Code: choose by your workflow

Both work on project code. Compare entry points, account setup, skills and MCP, then use the same small task to decide which fits your workflow.

Start with their shared capabilities

Codex and Claude Code offer terminal and editor workflows for reading projects, editing files and running development commands. Code generation or MCP support alone does not decide fit, and this guide assigns no performance ranking.

Check where you will start

Codex CLI documents signing in and starting from a project directory; its extension works beside open files and selections. Claude Code documents terminal and other entry points. Check account access, your editor and task permissions before choosing a starting point.

If a team already has working instructions or skills, check how the selected tool reads and configures them. Keep a workflow you can understand and review.

Concrete differences in access and entry points

Codex: local CLI and IDE work support ChatGPT sign-in or an OpenAI API key. Codex cloud requires ChatGPT sign-in, and API-key access has different limitations for some workspace or cloud features.

Claude Code: the terminal quickstart requires an eligible Claude subscription, a Claude Console account or a supported cloud-provider connection. Organization-provided access needs the corresponding setup; a registered account alone does not establish access everywhere.

Start with an access method and workflow you already have. If neither is available, check official access requirements before choosing; a comparison does not require opening two services first.

Project instruction files differ too

Codex uses AGENTS.md for persistent project guidance such as conventions and checks. Claude Code commonly uses CLAUDE.md. These are project instructions for an assistant, not models or plugins.

Current Claude Code documentation also supports reading existing AGENTS.md files through configuration, with scope and version conditions. Check the files and settings rather than rewriting all guidance just to switch tools.

Skills and MCP are not exclusive advantages

Both support repeatable instructions and external tools. Check a skill pack or MCP service’s format, configuration, authorization and current entry point. Working in one environment does not imply universal compatibility.

Use the same checkable task

This is a selection method, not an installation test or benchmark by this website. Check official guidance for models, access, costs and quotas.

  1. Give each tool the same task, files and expected result, such as a known bug fix with existing checks.
  2. Consider whether you can follow the approach, review edits and verify the result, not only the final answer.
  3. Choose a starting tool for your account, entry point and review habits. A small task need not require two extension setups.

Sources and ranking

This directory lists AI development plugins with public source code and a clear purpose. Browse by task, then open an introduction to learn what a project does, who it is for, and what you need to get started. It covers selected projects, not every AI plugin.

How do the three plugin types differ?

Tool connections (MCP) let AI assistants use external tools such as browsers, code repositories and databases. MCP is the protocol used for these connections.

Reusable skills give a compatible AI assistant instructions and guidance for a particular task.

Editor plugins run inside the software you use to write code, adding AI chat, completions or suggested edits.

How does “Most starred” work?

Projects are sorted by their total GitHub stars. Stars indicate interest; they are not downloads, active users or quality scores. A skill pack counts as one project. Daily star growth is not shown without historical records.

What does “Recently updated” mean?

It uses the most recent code push, which may not be a new release. The 7-day filter checks that project update time; ranking still uses all-time stars. Each introduction shows when its star count was collected. Failed updates keep the last saved data.

Introductions are based on public project documentation. These projects have not each been installed and tested here. Check the project documentation for setup, compatibility and service costs. Refreshing this list only reads saved data; it does not start collection.

UTC+8