Claude Code と Codex の主要機能を比較する｜MCP・Skills・メモリ・Subagents・Hooks・Plugins リファレンス A feature-by-feature comparison of Claude Code and OpenAI Codex CLI, covering MCP, memory,…

As AI coding agents mature, two terminal-native tools have emerged as the primary benchmarks for developers: Anthropic's Claude Code and OpenAI's Codex CLI. A recently published Qiita reference article breaks down their feature sets side by side, offering a useful map for teams evaluating which tool fits their workflow.

Both agents share the same core premise — conversational, iterative coding directly from the terminal — but they diverge significantly in how they handle extensibility. Claude Code is built around the Model Context Protocol (MCP), an open specification championed by Anthropic that standardizes how agents connect to external tools and data sources. MCP has attracted a growing ecosystem of third-party server implementations, making it a potential interoperability layer across multiple agent frameworks. Codex, by contrast, relies on a Plugins and Skills model that integrates more tightly with OpenAI's own platform conventions.

Memory architecture is another point of contrast. Claude Code uses file-based context management — most notably CLAUDE.md files placed at the project root — which are automatically ingested at session start. This approach keeps context transparent and version-controllable. Codex appears to lean toward more explicit API-level session memory, though its long-term persistence model continues to evolve.

On the orchestration side, Claude Code's Subagents feature allows parallel execution of multiple agent instances, enabling task decomposition across large codebases or monorepos. This is particularly relevant for ambitious refactoring tasks where a single context window would be insufficient. Hooks — a Claude Code-specific capability — let developers inject custom scripts at defined lifecycle events such as pre/post file save or command execution, opening natural integration points with CI pipelines and git workflows.

The broader competitive landscape adds useful context. Editor-integrated agents like Cursor, Windsurf, and GitHub Copilot Workspace are racing to add similar orchestration and memory capabilities, blurring the line between terminal-first and editor-first paradigms. If MCP gains wider adoption beyond Anthropic's own tooling, it could become a meaningful interoperability standard — allowing teams to swap underlying models while preserving tool integrations, much like LSP did for language tooling in editors.

For teams making a practical choice today, the decision likely comes down to ecosystem fit. Claude Code's MCP-centric design suits teams already investing in open, composable toolchains, while Codex may appeal to those deeper in the OpenAI platform. Either way, the feature parity between the two is narrowing rapidly, and the comparison article serves as a useful snapshot of where each tool stands in mid-2025.