AgentSkillsCN

claude-code-cli

使用Claude Code CLI构建和运行AI代理。当开发自主代理、多代理系统、CI/CD自动化或为程序化任务编写Claude脚本时，请使用此指南。涵盖认证、无头模式（-p）、JSON输出解析、工具限制、子代理和编排模式。

SKILL.md
--- frontmatter
name: claude-code-cli
description: Build and run AI agents using Claude Code CLI. Use when developing autonomous agents, multi-agent systems, CI/CD automation, or scripting Claude for programmatic tasks. Covers authentication, headless mode (-p), JSON output parsing, tool restrictions, subagents, and orchestration patterns.
license: MIT
allowed-tools:
  - Bash
  - Read
  - Write
  - Edit

Claude Code Agent Development

Build autonomous agents using Claude Code CLI. Language-agnostic patterns for agent orchestration and programmatic integration.

Quick Reference

TaskCommand
Basic agentclaude -p "prompt"
JSON outputclaude -p --output-format json "prompt"
Streamingclaude -p --output-format stream-json "prompt"
Autonomousclaude -p --dangerously-skip-permissions "prompt"
Budget limitclaude -p --max-budget-usd 5 "prompt"
Restrict toolsclaude -p --tools "Read,Edit" "prompt"
Structured outputclaude -p --output-format json --json-schema '{...}' "prompt"

Authentication

API Key

bash
export ANTHROPIC_API_KEY="sk-ant-..."

OAuth Token (Pro/Team Subscription)

bash
claude setup-token  # Interactive setup
export CLAUDE_CODE_OAUTH_TOKEN="your-token"

Running Agents

Headless Mode (-p)

bash
# Basic prompt
claude -p "Implement a REST API"

# Piped input
echo "Fix the bug" | claude -p

# Full autonomy (sandboxed environments only)
claude -p --dangerously-skip-permissions "Build the feature"

Output Formats

FormatFlagUse Case
textdefaultSimple scripts
json--output-format jsonProgrammatic parsing
stream-json--output-format stream-jsonReal-time UI

For detailed parsing examples: See references/response-parsing.md

Structured Output

bash
echo "What is 2+2?" | claude -p --output-format json \
  --json-schema '{"type":"object","properties":{"answer":{"type":"number"}},"required":["answer"]}'

Result: {"type": "result", "structured_output": {"answer": 4}, ...}

Tool Restrictions

bash
# Read-only agent
claude -p --tools "Read,Glob,Grep" "Analyze code"

# Specific bash commands only
claude -p --allowed-tools "Bash(npm:*) Bash(git:*)" "Build project"

# Deny dangerous operations
claude -p --disallowed-tools "Bash(rm:*)" "Clean up"

For project configuration: See references/configuration.md

Multi-Agent Patterns

Sequential Pipeline

bash
analysis=$(claude -p --output-format json "Analyze codebase")
claude -p "Implement based on: $analysis"

Parallel Agents

bash
claude -p "Review auth/" > auth.txt &
claude -p "Review api/" > api.txt &
wait
claude -p "Summarize: $(cat *.txt)"

Specialist Agents

bash
claude --agents '{"security": {"prompt": "You are a security expert"}}' \
  --agent security -p "Review auth"

For complete orchestration patterns: See references/orchestration.md

Session Management

bash
# Persistent session
claude --session-id "task-123" -p "Start feature"
claude --session-id "task-123" -c -p "Continue"

# Stateless
claude -p --no-session-persistence "One-off task"

Project Configuration

.claude/settings.json

json
{
  "permissions": {
    "allow": ["Bash(npm:*)", "Edit", "Read"],
    "deny": ["Bash(rm -rf:*)"]
  }
}

CLAUDE.md

Project instructions auto-loaded by Claude.

For CI/CD and MCP: See references/configuration.md

Parsing JSON Output

Quick jq Examples

bash
# Get result
claude -p --output-format json "prompt" | jq -r '.[-1].result'

# Check success
claude -p --output-format json "prompt" | jq '.[-1].is_error'

# Get cost
claude -p --output-format json "prompt" | jq '.[-1].total_cost_usd'

Message Types

TypeDescription
systemInit with tools, model, session
assistantClaude's response
userTool results
resultFinal status and cost

For complete parsing guide: See references/response-parsing.md

Best Practices

  1. •Budget limits: Always --max-budget-usd for autonomous agents
  2. •Tool restrictions: Minimal toolset with --tools
  3. •Structured output: --json-schema for reliable parsing
  4. •Sandboxing: --dangerously-skip-permissions only in isolated environments
  5. •Error handling: Check is_error and exit codes