Claude Code CLI Reference: Every Command and Flag

Key Takeaways#

• claude -p "query" is the workhorse for scripting and CI/CD — it runs a one-shot query in non-interactive print mode and exits.
• Permission modes (--permission-mode) let you trade safety for automation speed: use default for interactive work, acceptEdits or auto for semi-automated runs, and bypassPermissions only in sandboxed CI.
• Session flags (-c, -r, --session-id) let you continue or resume conversations, enabling multi-step workflows without repeating context.
• JSON output (--output-format json) gives you structured data including cost, duration, and turn count — essential for monitoring and alerting in production pipelines.
• Environment variables control provider backends (Bedrock, Vertex, Foundry), auto-updates, and API key injection — set them before any scripting.

---

Core Commands#

---

Session & Interaction Flags#

---

Model & Effort Flags#

The --effort flag directly controls how much reasoning the model applies. low is fast and cheap for simple tasks; max engages deep chain-of-thought for complex problems.

---

Permission Flags#

Warning: bypassPermissions and --dangerously-skip-permissions should only be used in fully sandboxed environments (Docker containers, CI runners with no sensitive access). Never use them on a local development machine with real credentials.

---

Non-Interactive Mode (-p) Deep Dive#

The -p / --print flag is the foundation for all scripting and automation with Claude Code.

Basic One-Shot Query#

claude -p "Explain the purpose of the config.yaml file in this repo"

Runs the query, prints the response to stdout, and exits with code 0.

Piped Input#

cat error.log | claude -p "Analyze this error log and identify the root cause"

Pipe any content into stdin. Claude Code reads it as context before processing the prompt.

Output Formats#

JSON Output Structure#

When using --output-format json, the output is a JSON object:

{
  "type": "result",
  "subtype": "success",
  "total_cost_usd": 0.0042,
  "duration_ms": 3210,
  "num_turns": 2,
  "result": "The root cause is a missing database connection...",
  "session_id": "abc123-def456"
}

Turn and Bare Control#

--max-turns is critical for CI pipelines — it prevents runaway conversations from burning through your API budget.

---

Key Environment Variables#

Provider variables are mutually exclusive — set only one of USE_BEDROCK, USE_VERTEX, or USE_FOUNDRY. If none are set, Claude Code uses the direct Anthropic API with ANTHROPIC_API_KEY.

---

Scripting Patterns#

Piping Files for Analysis#

cat src/auth/errors.go | claude -p "Explain each error type and suggest better messages"

CI Integration with Structured Output#

claude -p "Review the diff and flag any security concerns" \
  --output-format json \
  --max-turns 5 \
  --permission-mode bypassPermissions

Parse the JSON in your CI script:

RESULT=$(claude -p "..." --output-format json --max-turns 3)
COST=$(echo "$RESULT" | jq -r '.total_cost_usd')
DURATION=$(echo "$RESULT" | jq -r '.duration_ms')
TEXT=$(echo "$RESULT" | jq -r '.result')
echo "Cost: \${COST} | Duration: \${DURATION}ms"

Scheduled / Cron Tasks#

# Daily codebase health check (crontab)
0 6 * * * cd /home/user/project && claude -p "Summarize open TODOs and FIXMEs" --output-format json --max-turns 3 >> /var/log/claude-health.log

Multi-Step Workflows with Session IDs#

# Step 1: Analysis
SESSION=$(claude -p "Analyze the test failures" --output-format json | jq -r '.session_id')

# Step 2: Fix (continue the same session)
claude -p "Now fix the failures you identified" --session-id "$SESSION"

Error Handling#

RESULT=$(claude -p "query" --output-format json --max-turns 2)
SUBTYPE=$(echo "$RESULT" | jq -r '.subtype')

if [ "$SUBTYPE" = "error" ]; then
  echo "Claude Code failed" >&2
  echo "$RESULT" | jq -r '.result' >&2
  exit 1
fi

---

Quick Reference Cheat Sheet#

# Interactive
claude                              # start REPL
claude "fix the tests"              # start with prompt

# One-shot
claude -p "explain this file"
cat file | claude -p "analyze"

# Sessions
claude -c                           # continue last
claude -r                           # resume (pick)
claude --session-id ID              # resume specific

# Model
claude --model opus -p "hard task"
claude --effort max -p "deep think"

# Permissions
claude --permission-mode auto       # mostly auto
claude --allowedTools Read,Write    # restrict tools

# Output
claude -p --output-format json "q"
claude -p --max-turns 5 "q"

# CI
claude -p "q" --output-format json --max-turns 3 --permission-mode bypassPermissions