Terminal · Claude Code · Codex · Command line

Terminal Mastery with Claude Code and Codex: Commands, Flags, Shortcuts and Scripts

Claude Code and Codex are at their best in a terminal, because a terminal is where your files, Git, tests and scripts already live. This guide takes you from typing claude or codex to running them in scripts, in the background, in parallel worktrees and in CI. It lists the commands, flags, slash commands and keyboard shortcuts that matter, with copy-ready recipes. Checked against Claude Code 2.1.308 and Codex CLI 0.160.0; run --help on your own version, because flags change quickly.

1. Four ways to run an agent

2. Install, sign in, update, diagnose

3. Starting a session

# Claude Code
claude                              # interactive, in the current folder
claude "explain this project"       # interactive, with a first prompt
claude -c                           # continue the latest conversation here
claude -r "auth-refactor"           # resume by name or ID (no name: picker)
claude -n "my-feature"              # give the session a name
claude --permission-mode plan       # start in read-only plan mode
claude -w feature-auth              # work in an isolated git worktree

# Codex
codex                               # interactive
codex "fix the failing tests"       # interactive, with a first prompt
codex resume --last                 # continue the most recent session
codex resume <name-or-id>           # resume a specific one (no id: picker)
codex -C ../other-repo              # use another folder as the working root
codex --worktree                    # new managed git worktree for the session

4. Inside the session: input tricks and shortcuts

The five prefixes (Claude Code)

In Codex, type @ to mention a file or use /mention src/lib/api.ts, and attach images at launch with -i shot.png.

Claude Code keyboard shortcuts

For Codex, the interrupt and other keys are configurable. Run /keymap inside the CLI to see and change them, and /debug-config to see which config layer set a value.

5. Slash commands you will use daily

Claude Code

Codex CLI

Which commands appear depends on version and plan, so type / to see yours.

6. The flags that matter

Claude Code

Codex

7. Permissions and sandboxes from the terminal

The terminal makes it easy to run an agent with the keys to your machine. Decide the trust level on purpose.

Test a sandbox without the agent: codex sandbox runs a command inside a Codex-provided sandbox, so you can see what it blocks.

8. Scripting: pipes, JSON and CI

Claude Code one-shot recipes

# 1. Pipe content in
git diff --staged | claude -p "Write a one-line commit message"
cat error.log | claude -p "What is the root cause? Be brief."

# 2. Machine-readable result
claude -p "List the TODOs in src/" --output-format json | jq -r '.result'

# 3. Validated structure (the answer must match the schema)
claude -p "Summarise this repo" \
  --json-schema '{"type":"object","properties":{"summary":{"type":"string"}},"required":["summary"]}'

# 4. Guard rails for unattended use
claude -p "Fix the failing test" \
  --max-turns 8 --max-budget-usd 2 \
  --allowedTools "Read" "Edit" "Bash(npm test)" \
  --permission-mode acceptEdits

# 5. Fast, predictable start in CI (skips hooks, plugins, CLAUDE.md discovery)
claude --bare -p "Check for type errors in src/" \
  --append-system-prompt-file ./ci-rules.txt

# 6. Continue a previous run
claude -c -p "Now also update the tests"

The JSON output includes the final text in a result field along with session and cost details. stream-json emits events as they happen (it needs --verbose with -p), which suits live dashboards. --no-session-persistence avoids saving the run to disk.

Codex one-shot recipes

# 1. Run a task and exit
codex exec "Add input validation to the signup form"

# 2. Pipe content in (stdin is appended to your prompt)
git diff | codex exec "Summarise this change for a PR description"

# 3. Events as JSON lines, and the final answer to a file
codex exec --json "Find unused dependencies" | jq -c 'select(.type)'
codex exec -o answer.txt "Summarise the architecture"

# 4. A required answer shape
codex exec --output-schema ./schema.json "List modules and owners"

# 5. Read-only, no saved session, outside a git repo
codex exec -s read-only --ephemeral --skip-git-repo-check "Explain this folder"

# 6. Review without opening the UI
codex review --uncommitted          # staged, unstaged and untracked changes
codex review --base main            # against a base branch
codex review --commit <sha>        # one commit
codex review "Focus on security"    # custom instructions

# 7. Continue or fork a non-interactive session
codex exec resume --last "Now add tests"

If stdin is piped and a prompt is given, Codex appends stdin as a block after the prompt. With no prompt (or -), the whole prompt is read from stdin.

Loops and batches

# Review every changed file, one call each
git diff --name-only main | while read f; do
  claude -p "Review $f for bugs. Reply with 'OK' if none." --allowedTools "Read" < /dev/null
done

# Parallel fan-out, four at a time
ls src/*.kt | xargs -P4 -I{} sh -c 'claude -p "Add KDoc to {}" --permission-mode acceptEdits --max-turns 5'

CI and git hooks

• Authenticate with a secret: claude setup-token output or an API key for Claude, an API key for Codex. Never print it into logs.
• Make it deterministic: --bare for Claude, --ignore-user-config and --ephemeral for Codex.
• No questions in CI: deny instead of ask (--permission-prompts none for Claude, -a never for Codex).
• Gate on exit code, and keep a human approval before merging anything the agent wrote.
• Pre-commit example: git diff --staged | claude -p "Any secrets or debug prints? Answer NONE if clean." and fail the hook if the answer is not NONE. Keep such hooks fast and optional.

9. Sessions: resume, fork, background

Background sessions (Claude Code)

A background session keeps working after you close the terminal tab. A small supervisor process hosts them.

claude --bg "investigate the flaky checkout test"   # prints a session ID, returns at once
claude agents                  # live view of all background sessions
claude agents --json           # the same, for scripts
claude attach 7c5dcf5d         # reconnect to it in this terminal
claude logs 7c5dcf5d           # print recent output without attaching
claude stop 7c5dcf5d           # stop it (transcript and worktree are kept)
claude respawn 7c5dcf5d        # restart with the conversation intact
claude rm 7c5dcf5d             # remove it from the list
claude daemon status           # is the supervisor healthy?

Inside a running session, /background detaches it, and Ctrl+B pushes a single long command into the background. For Codex, codex agents browses agent sessions on the shared local server, and codex queue sends a message to an existing session.

10. Parallel work: worktrees and tmux

# Claude Code
claude -w feature-auth               # worktree at .claude/worktrees/feature-auth
claude -w feature-auth --tmux        # ...and open it in its own tmux session
claude -w "#123"                     # worktree branched from a pull request

# Codex
codex --worktree                     # managed git worktree for this session

# By hand (works for any tool)
git worktree add ../app-auth -b feature/auth
cd ../app-auth && claude

• tmux layout: one window, one pane per agent, plus a pane for git status. tmux new -s agents, Ctrl+b % to split.
• Claude agent teams can place each teammate in a tmux or iTerm2 pane: claude --teammate-mode auto (see the agents article).
• Clean up: git worktree list then git worktree remove ../app-auth once merged.

11. Configuration and troubleshooting

Where settings come from

When something is off

1. Check the basics: claude doctor / codex doctor, then claude auth status / codex login status.
2. Strip customisations: claude --safe-mode starts with CLAUDE.md, skills, plugins, hooks and MCP servers off. If the problem disappears, add things back one by one.
3. Read the logs: claude --debug='mcp,startup' or --debug-file /tmp/claude.log. In Codex, /debug-config shows which config layer set a value.
4. Check the layers: a project file can override your user file, and a managed policy overrides both.
5. Garbled screen? Ctrl+L in Claude Code. In Codex try --no-alt-screen.
6. Reinstall cleanly: claude install stable, or reinstall Codex with your package manager.

MCP and plugins from the shell

claude mcp                      # manage MCP servers (add, list, remove...)
claude mcp login sentry         # run a server's OAuth flow without the interactive panel
claude plugin install code-review@claude-plugins-official

codex mcp add / list / get / remove / login / logout
codex plugin add / list / remove
codex features list             # inspect feature flags

12. Make your terminal comfortable

• Shift+Enter for a new line: run /terminal-setup in Claude Code to install the keybinding in VS Code, Cursor, Alacritty or Zed. In Apple Terminal enable Option+Enter. Everywhere, a trailing \ then Enter also adds a line.
• Option as Meta (macOS): required for the Alt+ shortcuts. It is a profile setting in Terminal.app and iTerm2.
• Long prompts: Ctrl+G opens your $EDITOR. Set EDITOR in your shell profile.
• Vim keys in the prompt: /config then Editor mode. The prompt then supports NORMAL and INSERT modes, motions such as w, b, 0, $, and edits such as dd, cw, yy, p.
• Look and feel: /theme (including an auto light/dark option and colorblind-friendly themes), /statusline to show model, branch or cost in the footer, /color to tag a session.
• Flicker-free fullscreen: /tui fullscreen switches the renderer and keeps your conversation. /tui default goes back.
• Scrollback: for Codex --no-alt-screen keeps your terminal history; in Claude Code’s transcript view (Ctrl+O) you can search and open it in your editor.
• Accessibility: claude --ax-screen-reader gives plain, flat output for screen readers.
• Custom keys: /keybindings (Claude Code) or /keymap (Codex).

13. Aliases and shell recipes

Put these in ~/.zshrc or ~/.bashrc. They are examples; change the names and flags to taste.

# short names
alias cc='claude'
alias ccc='claude -c'                       # continue
alias ccp='claude --permission-mode plan'   # safe exploration
alias cx='codex'
alias cxr='codex resume --last'

# ask a question about the code, read-only, answer on stdout
ask() { claude -p "$*" --permission-mode plan; }

# commit message from the staged diff
cmsg() { git diff --staged | claude -p "Write a concise conventional commit message. Output only the message."; }

# explain the last failing command's output
why() { "$@" 2>&1 | tail -50 | claude -p "Why did this fail? Give the likely cause and the fix."; }

# review my branch against main with Codex
rv() { codex review --base "${1:-main}"; }

# new worktree + agent in one go
wt() { git worktree add "../$(basename "$PWD")-$1" -b "feature/$1" && cd "../$(basename "$PWD")-$1" && claude; }

14. Cheat sheet and checklist

Before you hand a task to an agent in the terminal

1. Are you in the right folder and on the right branch (git status)?
2. Is the trust level right: plan or read-only for exploring, narrow allow rules for editing, no bypass flags outside a sandbox?
3. Is the work in a worktree if other things are in progress?
4. For scripts: a turn cap, a budget cap, no prompts, and a human gate before merge?
5. After the run: git diff, run the tests, and read the change yourself.

Links and further reading

• Claude Code CLI reference · Interactive mode · Commands · Run Claude Code programmatically
• Codex CLI commands and flags
• On this site: Claude Code series · Codex series · Configuring AI Coding Agents · AI Engineering Quicksheet · Git quicksheet