Claude Code vs OpenAI Codex: Complete Guide — Installation, Commands & Real-World Examples

Goal of this article: Not "which one is better?" — but showing you, with real commands and examples, how to install, run, and use both tools effectively.

---

Why Should You Learn Both Tools Right Now?

Claude Code (Anthropic) and OpenAI Codex CLI (OpenAI) are terminal-based AI coding agents that emerged in 2025–2026. Unlike IDE autocomplete tools such as GitHub Copilot, both are autonomous agents that read your entire codebase, directly edit files, and run tests.

They belong to the same category but operate with different philosophies. Understanding those differences at the command level lets you make the right choice for each situation.

---

Claude Code: Core Features and Commands

How Do You Install Claude Code?

# Install (requires Node.js 18+)
npm install -g @anthropic-ai/claude-code

# Set your API key
export ANTHROPIC_API_KEY="sk-ant-..."

# Check version
claude --version

# Start an interactive session
claude

# Start with current directory as context
cd my-project && claude

What Execution Modes Are Available?

Mode	Command	Description
Interactive	claude	Interactive session (exit with /exit)
One-shot	claude "prompt"	Runs once and exits
Pipe mode	cat file | claude -p "..."	Reads from stdin, writes to stdout
Resume last	claude --continue	Resumes the most recent session
Resume by ID	claude --resume <id>	Returns to a specific session

# One-shot — complete a task in a single line
claude "Find and fix the JWT expiry bug in app/api/auth/route.ts"

# Pipe mode — use git diff for code review
git diff main | claude -p "Review these changes: security, performance, then readability"

# Analyze error logs
cat error.log | claude -p "Identify the root cause and suggest a fix"

Which Slash Commands Can You Use Inside a Session?

These commands are available within an interactive session (claude).

Command	Purpose
/init	Analyze the current project and auto-generate CLAUDE.md
/clear	Reset context (use when starting a new task)
/compact	Compress the conversation — saves tokens in long sessions
/cost	Show token usage and estimated cost for the current session
/model	Switch model (Haiku ↔ Sonnet ↔ Opus)
/memory	Manage Auto-Memory — view and toggle saved context (v2.1.59+)
/permissions	View and change allowed tool list
/help	Show all available commands
/btw [question]	Ask a quick side question without adding it to conversation history (v2.1.72+)
/vim	Toggle Vim key bindings

# Example usage inside a session
$ claude
> /init
✓ CLAUDE.md created — project structure analysis applied

> /cost
Session tokens: 12,400 input / 3,200 output | Estimated cost: $0.08

> /compact
✓ Conversation compressed — 40% context saved

> /btw What JWT library is being used in auth.ts right now?
↳ [Side question — not saved to history] Using jsonwebtoken 0.9.1

Key characteristics of /btw:

• Answers using the current context — cannot read new files or run commands
• The response is not added to conversation history, so it does not pollute context
• Can run in parallel even while Claude is working on a task
• Reuses the parent conversation's prompt cache, minimizing extra cost

Why Does the CLAUDE.md Configuration File Matter?

CLAUDE.md is a project-specific instruction file placed at the project root. Claude Code reads it automatically at session start and applies your coding rules, structure notes, and restrictions.

# CLAUDE.md example

## Stack
Next.js 16 App Router, TypeScript strict, Tailwind CSS, Drizzle ORM + Neon Postgres

## Coding Rules
- Functional components only (no class components)
- No `any` types — use unknown + type guards instead
- Component files: PascalCase, util files: camelCase
- All new API routes must apply rate-limit middleware

## Testing
- Test runner: `node --experimental-strip-types --test`
- Write unit test files alongside any new utility functions

## Restrictions
- Never modify `.env` files
- Do not commit with leftover `console.log` statements
- Do not cast `any` to bypass type errors

Without CLAUDE.md: You must explain project rules at the start of every session. With CLAUDE.md: Define project rules once — they apply automatically to every session.

How Do Skills Package Reusable Workflows? (v2.1+)

Skills let you package frequently used tasks into SKILL.md files for reuse. Place them under .claude/skills/ or ~/.claude/skills/ and they hot-reload immediately. From Claude Code 2.1 onward, they appear integrated with slash commands when you type /.

# .claude/skills/code-review.md
---
name: code-review
description: Review PR diff in order of security, performance, readability
user-invocable: true
---

Read `git diff HEAD~1` and review in this order:
1. Security vulnerabilities (injection, missing auth, exposed secrets)
2. Performance issues (N+1 queries, unnecessary re-renders)
3. Readability and maintainability

$ claude
> /code-review        # Direct invocation inside a session (user-invocable: true)
> /commit             # Analyze staged changes and generate a commit message

Two invocation modes for Skills:

Mode	Frontmatter	Trigger
User-invocable	user-invocable: true	Call directly with /skill-name
Auto-invocation	user-invocable: false	Claude reads description and triggers automatically

Adding context: fork runs the skill in a sub-agent isolated from the current session — protecting session tokens during long analysis tasks.

How Does Auto-Memory Preserve Context Across Sessions? (v2.1.59+)

Auto-Memory automatically saves what Claude learns during a session — build commands, debugging patterns, architecture notes, coding style preferences — and reuses that knowledge in future sessions. It is enabled by default.

$ claude
> /memory

📁 ~/.claude/projects/my-project/memory/MEMORY.md
  - Build: pnpm build (Next.js 16)
  - Test: node --experimental-strip-types --test
  - Note: Never modify .env files (recorded: 2026-03-15)
  [auto-memory toggle: ON]

CLAUDE.md vs Auto-Memory — how they differ:

	CLAUDE.md	Auto-Memory
Written by	Developer (manual)	Claude (automatic)
Content	Coding rules, restrictions	Build commands, debugging insights
Sharing scope	Entire team (git-trackable)	Personal session context
Control	Manual editing	Toggle with /memory command

What Are Hooks and How Do They Automate Your Workflow? (2026)

Hooks connect shell commands to Claude Code events. Configure them in .claude/settings.json.

// .claude/settings.json
{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{ "type": "shell", "command": "pnpm run lint --fix" }]
    }],
    "Stop": [{
      "hooks": [{ "type": "shell", "command": "echo 'Session ended' >> .claude/session.log" }]
    }]
  }
}

Key hook events: PreToolUse (before tool runs) | PostToolUse (after tool runs) | PreCompact (before compression) | Stop (session end)

→ This enables CI/CD-level automation: run the linter automatically every time AI edits a file, trigger type checks before commits, and more. Claude Code supports 12 hook events in total.

Real-World Command Examples

1. Autonomous bug-fix loop

claude "Find the JWT expiry bug in app/api/auth/route.ts, fix it,
then add 3 failing-case tests and keep running until all pass"

Claude Code autonomously cycles through: read files → identify bug → fix → run tests → re-fix on failure — until done, without human intervention.

2. AI-generated commit message

git add -p          # Stage your changes
claude commit       # Analyze staged changes and generate a commit message

3. Large-scale refactoring

claude "Convert all Promise chains (.then/.catch) in the lib/ directory
to async/await. Verify TypeScript types remain intact throughout."

4. Codebase onboarding

claude "Explain the authentication flow in this project step-by-step
for a developer seeing it for the first time, with file paths."

What Are the Key Flags?

# Restrict to read-only tools — analysis with no file modifications
claude --allowedTools "Read,Grep,Glob" "Scan for security vulnerabilities only"

# Autonomous execution without confirmations (use with caution)
claude --dangerouslySkipPermissions "Generate all test files in bulk"

# Specify a model
claude --model claude-opus-4-6 "Help me design a complex system"
claude --model claude-haiku-4-5-20251001 "Write a simple utility function"

---

OpenAI Codex: Core Features and Commands

How Do You Install OpenAI Codex CLI?

# Install (MIT open source, Node.js 22+ recommended)
npm install -g @openai/codex

# Set your API key
export OPENAI_API_KEY="sk-..."

# Check version
codex --version

# Start an interactive session
codex

# One-shot execution
codex "Add accessibility aria attributes to components/Button.tsx"

What Is Approval Mode?

This is Codex's key differentiator. You can set the user approval level for execution in 3 stages.

# suggest mode (default): confirms before every change
codex --approval-mode suggest "Refactor this module"

# auto-edit mode: file edits are auto-approved, shell commands require confirmation
codex --approval-mode auto-edit "Fix all TypeScript errors"

# full mode: file edits and shell commands both run automatically (use carefully)
codex --approval-mode full "Run tests automatically and fix failures"

Mode	File edits	Shell commands	Recommended for
suggest	Confirm first	Confirm first	Unfamiliar codebases, production
auto-edit	Automatic	Confirm first	Everyday development (recommended default)
full	Automatic	Automatic	Test environments, CI/CD

Which Workflows Does Codex Suit Best?

Codex uses a sandboxed execution environment by default. Shell commands run with restricted network access and isolated disk writes. This makes it a strong fit for security-sensitive enterprise environments.

# Multimodal — pass a screenshot
codex "Write a test that reproduces the UI bug in this screenshot" \
  --image ./screenshots/bug-report.png

# Quiet mode — use in scripts and pipes
cat legacy-code.js | codex -q "Convert this code to TypeScript"

Real-World Command Examples

1. Function refactoring

codex --approval-mode auto-edit \
  "Consolidate the 3 useState hooks in components/Header.tsx into a single useReducer"

2. Automated test generation

codex "Write Vitest unit tests for all public functions in lib/geo-probe.ts.
Include edge cases."

3. Generate a GitHub Actions workflow

codex "Write a CI workflow .github/workflows/ci.yml for this project
(Next.js, pnpm, Node 22). Order: lint → typecheck → test → build"

4. Code explanation (pipe mode)

cat lib/geo-check-scoring.ts | codex -q \
  "Explain this scoring logic in plain English without inline comments"

Codex Project Configuration (AGENTS.md)

AGENTS.md is Codex's equivalent of Claude Code's CLAUDE.md — a project instruction file for the agent.

# AGENTS.md

## Project
Next.js 16, TypeScript strict, pnpm

## Rules
- Do not change existing code style
- Use Vitest for all tests
- Write PR descriptions in English

How Do Agent Skills Package Workflows?

Codex also packages repetitive tasks using SKILL.md. They work identically across the CLI, IDE extension, and the Codex app. A community skills index is publicly available at github.com/openai/skills.

# .codex/skills/test-coverage.md
---
name: test-coverage
description: Auto-generate Vitest tests for files with low coverage
---

Find files in lib/ that have no tests or coverage below 70%,
then write Vitest unit tests including edge cases.

codex /test-coverage      # Invoke the skill directly

Are Multi-Agent Workflows Supported?

Multi-agent workflows are well-suited for large tasks that can be parallelized. Enable them with /experimental.

# Enable multi-agent inside a session
> /experimental

# CSV-based parallel fan-out — process multiple modules simultaneously via sub-agents
codex "Refactor each module in tasks.csv using parallel agents"

Using spawn_agents_on_csv, you can fan out tasks from a CSV list and track progress and ETA on a single screen. Sub-agents can be given nicknames so you can tell which agent is handling which task.

SWE-bench Verified (codex-1): 72.1% (1 attempt) / 83.8% (8 attempts)

What New Features Did 2026 Bring to Codex?

Feature	Details
GPT-5.4 support	Specify --model gpt-5.4 or use it as the default model
Codex-Spark	Research preview (ChatGPT Pro only), 128K context window
Fast mode on by default	Live Fast/Standard mode indicator in TUI header
Voice input	Hold spacebar → record and transcribe (experimental: features.voice_transcription = true)
Windows native	PowerShell + Windows Sandbox (no WSL required)
/theme picker	Live preview of light/dark theme inside a session
TUI code highlighting	Syntax highlighting for fenced code blocks and diffs

---

How Do Both Tools Handle the Same Task Differently?

Scenario 1: Bug Fix + Autonomous Test Loop

Claude Code:

claude "Fix the session expiry bug in app/api/auth/route.ts,
add 3 failing-case tests, and keep iterating until all pass"

→ Autonomous loop: reads files → identifies bug → fixes → runs tests → re-fixes on failure. Runs to completion with no human input.

OpenAI Codex:

# Step 1: Fix the bug (you can review the diff for each file change)
codex --approval-mode auto-edit \
  "Fix session expiry bug in app/api/auth/route.ts"

# Step 2: Add tests
codex --approval-mode auto-edit \
  "Add 3 failing-case tests to auth.test.ts"

# Step 3: Run tests (shell commands require confirmation)
codex --approval-mode full "pnpm test auth"

→ Step-by-step control: review and approve/reject each file diff. More control at each stage.

Scenario 2: Code Review

Claude Code:

git diff main | claude -p "Review this PR.
List issues in order: security vulnerabilities, performance, readability."

OpenAI Codex:

git diff main | codex -q "Review this PR.
List issues in order: security vulnerabilities, performance, readability."

Both tools work equally well in pipe mode for this workflow.

How Do the Commands Compare Side by Side?

Task	Claude Code	OpenAI Codex
Autonomous bug fix	claude "find and fix, run until tests pass"	codex --approval-mode full "..."
Read-only exploration	claude --allowedTools "Read,Grep,Glob"	codex --approval-mode suggest
Pipe input	cat f | claude -p "..."	cat f | codex -q "..."
AI commit message	claude commit	Not supported (manual)
Resume session	claude --resume <id>	Session history (codex --history)
Project instructions	CLAUDE.md	AGENTS.md
Skills	/skill-name (user-invocable + auto-invocation)	/skill-name (Agent Skills)
Session memory	Auto-Memory (auto-save) + /memory	None (AGENTS.md manual only)
Hook automation	PostToolUse → auto lint/test	Not supported
Multi-agent	context: fork sub-agent fork	/experimental (experimental)
Voice input	Not supported	Experimental (hold spacebar)
Model selection	--model claude-sonnet-4-6	--model gpt-5.4 or codex-spark

---

Which Tool Should You Choose and When?

Situation	Recommendation	Reason
Understanding a large, unfamiliar codebase	Claude Code	1M context window, autonomous multi-file editing
Autonomous bug-fix loop	Claude Code	Runs to completion without interruption
Reviewing diffs before every change	Codex	Step-by-step approval with --approval-mode
Security-isolated execution environment	Codex	Sandboxed by default, network restricted
Open-source tool or self-modification needed	Codex	MIT license
Automated AI commit messages	Claude Code	claude commit built in
Image / screenshot input	Codex	Multimodal --image support
Team already using Claude API	Claude Code	Same API key, unified cost management
Accumulating context across sessions	Claude Code	Auto-Memory enabled by default
CI/CD automation pipeline	Claude Code	Hooks connect lint/test automatically
Large parallel code tasks	Codex	Multi-agent spawn_agents_on_csv
Windows native environment	Codex	PowerShell sandbox (no WSL needed)

---

Quick Reference Summary

	Claude Code	OpenAI Codex CLI
Install	npm i -g @anthropic-ai/claude-code	npm i -g @openai/codex
API key	ANTHROPIC_API_KEY	OPENAI_API_KEY
Default model	Claude Sonnet 4.6	GPT-5.4 / Codex-Spark
Execution control	--allowedTools / --dangerouslySkipPermissions	--approval-mode suggest/auto-edit/full
Project instructions	CLAUDE.md	AGENTS.md
Skills	SKILL.md (user-invocable + auto-invocation)	SKILL.md (Agent Skills)
Auto-Memory	Enabled by default (v2.1.59+)	Not available
Hooks	12 events (PostToolUse, etc.)	Not available
Multi-agent	context: fork	/experimental
Commit support	claude commit	Not supported
Sandbox	Optional	Default
Open source	No	Yes (MIT)
Voice input	Not supported	Experimental
Pricing model	Anthropic API usage	OpenAI API usage

---

FAQ

Related Terms (Glossary)

• AI Agent
• LLM (Large Language Model)
• Context Window
• Prompt Engineering

---

Further Reading

• This Week in AI Signals: The Fallout from the "90% AI-Written Code" Prediction
• Developer Survival Strategies in the AI Era: 5 Transitions to Start Right Now
• When 90% of Code Is Written by AI: How Will Developers Survive?
• The Evolution of AI Coding Assistants: From Autocomplete to Autonomous Agents

---

About This Article

This article was written in March 2026 based on the Anthropic Claude Code official documentation and the OpenAI Codex CLI GitHub README. Both tools update rapidly — new features will be reflected as they are released.

---

References

• Anthropic: Claude Code Official Documentation
• Claude Code Skills Official Documentation
• Claude Code Memory Official Documentation
• Claude Code Changelog
• OpenAI Codex CLI GitHub
• OpenAI Codex Agent Skills Documentation
• OpenAI Codex Multi-agent Guide
• Claude Code CLI Reference