claude-cli-runner

Claude CLI Runner

Use this skill when the user needs to run Claude models from Python code WITHOUT API keys — via the claude CLI binary that uses Claude Code's built-in authentication.

When to Use

• User asks to "run Claude from Python" or "call Claude without API key"
• User needs to process text with Claude in a backend/script
• User wants to integrate Claude into a project that doesn't have API keys configured
• User mentions "claude CLI" or "claude binary"

Module Location

${WORKSPACE}/tools/claude_cli.py

Quick Usage

import sys
sys.path.insert(0, "${WORKSPACE}/tools")
from claude_cli import claude, claude_async, claude_json, claude_stream, validate_response

# Simple call
result = claude("Fix spelling: Привет мр")

# With system prompt and model
result = claude(
    "Review this code for bugs",
    system="You are a senior developer",
    model="claude-opus-4-6",
)

# Async
result = await claude_async("Translate to English: ...")

# JSON output
data = claude_json("List 3 colors as JSON array")

# Streaming
for chunk in claude_stream("Write a story"):
    print(chunk, end="")

# Validate LLM response
ok, cleaned = validate_response(original_text, llm_response)

Available Models

• claude-opus-4-6 — most capable
• claude-sonnet-4-5 — balanced (default)
• claude-haiku-4-5 — fastest, cheapest

Requirements

Claude CLI must be installed: npm install -g @anthropic-ai/claude-code

Or set CLAUDE_CLI_PATH env var to the binary path.

On Server (your-server)

ssh your-server "which claude"  # verify installation
ssh your-server "claude -p --model claude-sonnet-4-5 'Hello'"  # test