claude-launcher: Model Freedom for Claude Code

Two days ago, two major open source models dropped targeting Claude Code compatibility. GLM-4.7 and MiniMax M2.1. Both MIT licensed. Both 7-10x cheaper than Claude. Both available through OpenRouter.

The missing piece: actually using them without friction.

The Problem

Switching Claude Code to OpenRouter requires setting multiple environment variables:

export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="sk-or-..."
export ANTHROPIC_MODEL="z-ai/glm-4.7"

Every time you want to switch models, you’re editing shell configs or typing env vars. Want to try GLM for a session, then switch back to Claude? Manual process. Want different models for different task types? More env vars.

This friction matters. The whole point of OpenRouter’s model marketplace is experimentation. If switching is painful, you won’t switch.

claude-launcher

A CLI wrapper that handles the backend configuration for you:

npm install -g claude-launcher

First run prompts for backend selection and model choice. After that:

claude-launcher              # launches with saved settings
claude-launcher -m           # pick a different model
claude-launcher -a           # force Anthropic backend
claude-launcher -o           # force OpenRouter backend
claude-launcher -- --resume  # pass args through to claude

Features

• Interactive model picker: Searchable list of all OpenRouter models that support tool calling. Shows context length and pricing inline.

• Exacto variants: Models with :exacto variants (fine-tuned for tool use) are flagged and auto-selected when available.

• Role-specific models: Claude Code uses different models for sonnet/opus/haiku tasks. Configure each separately. Use a cheap model for quick tasks, expensive for complex ones.

• New model alerts: Notifies when new agentic models appear on OpenRouter since your last session.

• Zero lock-in: Switch back to standard Anthropic backend anytime with -a.

Why This Matters

The open source agentic moment created options. GLM-4.7 hits 73.8% on SWE-bench at 7x cheaper. MiniMax M2.1 reaches 72.5% at 10x cheaper. Both target Claude Code’s tool-calling patterns specifically.

But options only matter if you can actually use them. claude-launcher removes the friction between “I want to try this model” and actually trying it.

The framework stays the same. Only the model changes.

— The real unlock

This is the promise of Claude Code’s OpenRouter integration. claude-launcher just makes it one command instead of five.

Try It

npm install -g claude-launcher
claude-launcher

Source on GitHub. MIT licensed. Built with Bun.