How Claude Code and Codex differ
Claude Code is built by Anthropic. The docs list the terminal, VS Code and JetBrains extensions, a desktop app and the web, all sharing the same CLAUDE.md files, settings and MCP servers.
Codex is built by OpenAI. Its docs cover a CLI and an IDE extension. Note that developers.openai.com/codex now redirects to learn.chatgpt.com/docs, so older tutorials point there.
As of October 24, 2026, a feature checklist will not separate them: both have skills, subagents, hooks, plugins and MCP. What matters is where files go and what survives a move between the two.
Comparison table
FeatureClaude CodeCodexProject rulesCLAUDE.md, .claude/rules/*.md with paths; AGENTS.md since v2.1.277, only when no CLAUDE.md existsAGENTS.md and AGENTS.override.md from repo root down to the current folder, 32 KiB cap by defaultSkills.claude/skills/<name>/SKILL.md, ~/.claude/skills/.agents/skills/, ~/.agents/skills/, /etc/codex/skillsInvoking a skill/name or automatic by description$name, /skills or automatic by descriptionInstalling a skill/plugin install, npx skills add <repo> -a claude-code, gh skill install$skill-installer <name>, npx skills add <repo> -a codexSubagents.claude/agents/*.md: name and description in frontmatter, body is the system prompt.codex/agents/*.toml: name, description, developer_instructionsHookssettings.json, 33 events, handler types command, http, mcp_tool, prompt, agenthooks.json or [hooks] in config.toml, 12 events, must be trusted via /hooksPlugins.claude-plugin/plugin.json, marketplaces via /plugin marketplace add/plugins browser, marketplaces synced from GitHubMCP.mcp.json (JSON, mcpServers key), claude mcp add~/.codex/config.toml, [mcp_servers.<name>] tables, codex mcp addSlash commands.claude/commands merged into skills, old files still workcustom prompts in ~/.codex/prompts are deprecated in favor of skillsSources: Claude Code memory, Claude Code hooks, Codex AGENTS.md, Codex hooks, Codex subagents.
What carries over unchanged
Skills mostly do. Both read the open Agent Skills format. The catch is the path. Codex looks in .agents/skills, and the Claude Code docs do not mention that directory, so a skill placed only there is invisible to Claude Code. Install with npx skills add for both agents, or keep a copy in .claude/skills.
The second catch is frontmatter. Claude Code understands extra fields such as context: fork, hooks and model; Codex ignores them. For a skill meant to run in both, stick to the six spec fields. See how to write a SKILL.md.
Rules carry over through a bridge. Codex reads AGENTS.md natively. Claude Code skips AGENTS.md when a CLAUDE.md exists, so you add a single line, @AGENTS.md, to CLAUDE.md. The loading modes are covered in AGENTS.md vs CLAUDE.md.
Subagents and MCP configs do not. Markdown versus TOML, JSON versus TOML. Convert by hand or with rulesync or ruler.
Where the difference shows up day to day
Hooks first. Claude Code has 33 events and five handler types, and exit code 2 blocks the action. The Codex docs list 12 events, including PermissionRequest and PostCompact, and every non-managed hook needs explicit approval. Either one can block a dangerous shell command. Codex hooks can run scripts and MCP tools; an http handler type that posts to an endpoint on its own is documented only for Claude Code.
Then the skill listing budget. Claude Code reserves 1% of the context window for the list and truncates each description at 1,536 characters. Codex takes 2% of the window, or 8,000 characters when the window size is unknown, and shortens descriptions when space runs out. With a hundred skills installed, both start missing the right one.
Which one fits your setup
- Mixed-tool team with an existing AGENTS.md: Codex reads it as is; Claude Code needs the bridge line.
- Hook-driven automation (format after edit, block
rm -rf, pre-commit checks): Claude Code has more events, more handler types and more detailed docs. - Ready-made plugins: Anthropic runs two marketplaces: the official one and a community one where every plugin passed an automated security scan. See best Claude Code plugins.
- Solo project: pick the agent whose model you prefer. Formats will not lock you in as long as rules live in AGENTS.md and skills stay within the six standard fields.
Browse Claude Code skills and Codex skills.
