Claude Code CLI Provider

Use your Anthropic Max subscription to power Wunderland agents — no API key needed. The claude-code-cli provider invokes your locally-installed Claude Code CLI as a subprocess.

Officially Supported

Anthropic explicitly supports programmatic use of Claude Code via the -p flag. Their headless mode documentation provides examples for scripted calls, CI/CD pipelines, and automation. This is the intended use case.

Prerequisites

1. Claude Code CLI installed on your machine
2. Anthropic account logged in (Free, Pro, or Max plan)

Install Claude Code

npm install -g @anthropic-ai/claude-code

Or download from claude.ai/download.

Log in

claude

Follow the browser-based login flow. Once authenticated, Claude Code caches your credentials locally.

Setup with Wunderland

Interactive login

wunderland login

Select "Claude Code CLI (Max subscription)" from the menu. Wunderland will verify your installation and authentication automatically:

  Claude Code CLI Provider
  ────────────────────────
  ✓ CLI installed    /usr/local/bin/claude (v1.5.0)
  ✓ Authenticated
  ✓ Model available  claude-sonnet-4-20250514

  Ready to use! No API key needed.

Direct provider flag

wunderland login --provider claude-code-cli

Manual config

~/.wunderland/config.json

{
  "llmProvider": "claude-code-cli",
  "llmModel": "claude-sonnet-4-20250514"
}

Or per-agent:

agent.config.json

{
  "llmProvider": "claude-code-cli",
  "llmModel": "claude-opus-4-20250514"
}

Available Models

Model	Best For	Context	Output Limit
claude-sonnet-4-20250514 (default)	Most tasks — balanced speed and quality	200K	16K
claude-opus-4-20250514	Complex reasoning, deep analysis	200K	32K
claude-haiku-4-5-20251001	Fast, lightweight tasks	200K	8K

All models are $0 per token — your subscription handles billing.

Override the model:

wunderland chat --provider claude-code-cli --model claude-opus-4-20250514

Or via environment variable:

export CLAUDE_CODE_MODEL=claude-opus-4-20250514

How It Works

1. Wunderland spawns claude --bare -p --output-format json as a subprocess
2. Your conversation is piped via stdin
3. Claude Code uses your local authentication (no tokens leave your machine)
4. The JSON response is parsed back into AgentOS's standard format

The --bare flag strips Claude Code's own plugins, hooks, and MCP servers for clean, reproducible output. --max-turns 1 ensures a single completion per call (no agentic tool loops).

Tool Calling

When Wunderland needs the model to call tools, it uses Claude Code's --json-schema flag to enforce structured output. The model responds with either text or a structured tool call array. This gives reliable tool calling without any custom parsing.

Troubleshooting

"Claude Code CLI is not installed"

npm install -g @anthropic-ai/claude-code
# or
brew install claude-code  # macOS

"Not authenticated"

claude  # opens browser login flow

"Timed out"

Claude Code didn't respond within 2 minutes. Try:

• Using a faster model: claude-haiku-4-5-20251001
• Checking your internet connection
• Running claude --bare -p "test" manually to diagnose

"Rate limited"

Your Max plan's usage limits have been reached. Wait a few minutes and try again. Limits assume ordinary individual usage — running large automated pipelines may trigger enforcement.

vs. Anthropic API Key Provider

Feature	claude-code-cli	anthropic
Auth	Claude Code login (subscription)	ANTHROPIC_API_KEY (Console)
Cost	$0 per token (subscription)	Pay-per-token
Setup	Install CLI + login	Get API key from Console
Rate limits	Subscription limits	API rate limits
Best for	Personal use, development	Production, high-volume

Use claude-code-cli for personal development and experimentation. Use anthropic for production deployments where you need predictable billing and higher rate limits.