Openclaw Claude Code Skill

Claude Code Skill 🤖

Control Claude Code via MCP (Model Context Protocol). This CLI provides agent-level programmatic access to Claude Code — persistent sessions, effort control, context management, model switching, and agent teams.

Built for OpenClaw agents that need to drive Claude Code as a coding backend.

What's New in v1.2 🚀

• Cost Tracking — session-cost with full token/price breakdown (Claude, Gemini, GPT pricing)
• Session Branching — session-branch to fork + change model/effort in one step
• Hooks System — session-hooks with webhook callbacks for tool errors, context high, stop events
• Model Aliases — built-in aliases (opus/sonnet/haiku/gemini-flash/gemini-pro) + custom --model-overrides
• Config Files — --config agent.json to load session presets from file
• Cached Token Tracking — separate pricing for prompt cache hits

v1.1

• Effort Control — --effort low/medium/high/max at session and per-message level
• Ultrathink — --ultrathink flag for deep reasoning on complex tasks
• Plan Mode — --plan flag to have Claude create a plan before executing
• Context Management — session-compact to reclaim context, session-context to inspect usage
• Model Switching — session-model to switch models mid-session (with auto-resume)
• Auto-Resume — --auto-resume to restart stopped sessions transparently
• NDJSON Streaming — --ndjson for machine-readable streaming output
• API Timeout Control — configurable per-request timeouts with AbortController

Features

• 🔌 MCP Protocol — Direct access to all Claude Code tools
• 💾 Persistent Sessions — Maintain context across multiple interactions
• 🧠 Effort Control — low/medium/high/max effort levels + ultrathink
• 📋 Plan Mode — Claude creates a plan before executing
• 🔄 Context Management — Compact sessions, inspect token usage
• 🤝 Agent Teams — Deploy multiple specialized agents
• 🔧 Tool Control — Fine-grained control over which tools are available
• 📊 Budget Limits — Set spending caps on API usage
• 🌐 Multi-Model — Use any model via proxy (Gemini, GPT, etc.)
• 📡 Streaming — Real-time SSE streaming with NDJSON support

Installation

git clone https://github.com/Enderfga/openclaw-claude-code-skill.git
cd openclaw-claude-code-skill
npm install
npm run build
npm link  # optional: make CLI globally available

Requirements

• Node.js 18+
• Backend API server running (see Configuration)
• Claude Code CLI installed

Configuration

# Default: http://127.0.0.1:18795
export BACKEND_API_URL="http://your-server:port"

Quick Start

# Start a session with high effort
claude-code-skill session-start myproject -d ~/project \
  --permission-mode acceptEdits \
  --allowed-tools "Bash,Read,Edit,Write,Glob,Grep" \
  --effort high

# Send a task with streaming
claude-code-skill session-send myproject "Refactor the auth module" --stream

# Need deep reasoning? Use ultrathink
claude-code-skill session-send myproject "Design a new caching layer" --stream --ultrathink

# Want a plan first?
claude-code-skill session-send myproject "Add rate limiting" --stream --plan

# Running low on context? Compact it
claude-code-skill session-compact myproject

# Check context usage
claude-code-skill session-context myproject

# Switch model mid-session
claude-code-skill session-model myproject sonnet

# Change effort level
claude-code-skill session-effort myproject max

# Done
claude-code-skill session-stop myproject

Command Reference

Connection & Tools

claude-code-skill connect          # Connect to MCP server
claude-code-skill disconnect       # Disconnect
claude-code-skill status           # Check connection status
claude-code-skill tools            # List available tools

Direct Tool Calls

claude-code-skill bash "npm test"
claude-code-skill read /path/to/file.ts
claude-code-skill glob "**/*.ts" -p ~/project
claude-code-skill grep "TODO" -p ~/project -c
claude-code-skill call Write -a '{"file_path":"/tmp/test.txt","content":"Hello"}'
claude-code-skill batch-read "src/**/*.ts" "tests/**/*.test.ts" -p ~/project

Persistent Sessions

Starting Sessions

# Basic
claude-code-skill session-start myproject -d ~/project

# With effort and model
claude-code-skill session-start myproject -d ~/project \
  --model claude-opus-4-5 \
  --effort high

# With proxy (multi-model support)
claude-code-skill session-start gemini-dev -d ~/project \
  --model gemini-2.0-flash \
  --base-url http://127.0.0.1:8082

# Full options
claude-code-skill session-start advanced -d ~/project \
  --permission-mode plan \
  --allowed-tools "Bash,Read,Edit,Write" \
  --disallowed-tools "Task" \
  --max-budget 5.00 \
  --effort high \
  --append-system-prompt "Always write tests" \
  --add-dir "/tmp,/var/log"

Sending Messages

# Basic (blocks until complete)
claude-code-skill session-send myproject "Write unit tests"

# Streaming
claude-code-skill session-send myproject "Refactor this" --stream

# With effort override (per-message)
claude-code-skill session-send myproject "Quick fix" --effort low
claude-code-skill session-send myproject "Complex redesign" --effort max

# Ultrathink (shorthand for --effort high)
claude-code-skill session-send myproject "Analyze this architecture" --ultrathink

# Plan mode
claude-code-skill session-send myproject "Add caching" --plan

# Auto-resume stopped sessions
claude-code-skill session-send myproject "Continue" --auto-resume

# NDJSON output (for programmatic consumption)
claude-code-skill session-send myproject "Run tests" --stream --ndjson

# Custom timeout
claude-code-skill session-send myproject "Long task" -t 900000

Effort & Model Control

# Set effort level (persists across messages)
claude-code-skill session-effort myproject low      # Minimal thinking
claude-code-skill session-effort myproject medium   # Balanced (Opus 4.6 default)
claude-code-skill session-effort myproject high     # Deep thinking
claude-code-skill session-effort myproject max      # Maximum, no token limit (Opus 4.6 only)
claude-code-skill session-effort myproject auto     # Reset to default

# Switch model mid-session
claude-code-skill session-model myproject opus
claude-code-skill session-model myproject sonnet

Context Management

# Compact session to reclaim context window
claude-code-skill session-compact myproject

# Compact with custom summary
claude-code-skill session-compact myproject --summary "Finished auth, now on tests"

# Check context usage
claude-code-skill session-context myproject
# → Tokens used: 45231, Usage: 22.6%, Suggestions: [...]

Session Management

claude-code-skill session-list                    # List active sessions
claude-code-skill session-status myproject        # Detailed status
claude-code-skill session-history myproject -n 50 # View history
claude-code-skill session-pause myproject         # Pause
claude-code-skill session-resume-paused myproject # Resume
claude-code-skill session-fork myproject exp      # Fork session
claude-code-skill session-restart myproject       # Restart failed session
claude-code-skill session-stop myproject          # Stop

Session Search

claude-code-skill sessions -n 20              # List recent sessions
claude-code-skill session-search "bug fix"    # Search by query
claude-code-skill session-search --project ~/myapp
claude-code-skill session-search --since "2h"
claude-code-skill resume <session-id> "Continue" -d ~/project

Permission Modes