glm-switcher

A small switcher for running Claude Code against Z.AI (GLM) models or your LiteLLM proxy using Anthropic-compatible settings.

This is an English rewrite and clean-room implementation inspired by claude-code-glm-switcher by Shor73.

✨ What it does

• Adds launchers that set Anthropic-compatible env vars for Z.AI or LiteLLM
• Provides a menu to pick native Claude, GLM variants, or LiteLLM
• Reads API keys from the environment rather than storing them in scripts

✅ Prerequisites

• macOS or Linux
• claude CLI installed and available on PATH
• A Z.AI API key for direct GLM launchers, or a LiteLLM proxy URL and key for LiteLLM launchers

🚀 Quick start

1. Run the installer. It can optionally add your Z.AI key to your shell config:

./install.sh

2. Use the commands:

claude            # native Claude
claude-glm        # direct Z.AI, GLM-5.3
claude-glm-53     # same as claude-glm
claude-glm-53-flash # direct Z.AI, GLM-5.3-Flash
claude-litellm    # LiteLLM proxy with your configured model
claude-switch     # menu

🔀 How the mapping works

The direct GLM scripts set Anthropic-compatible environment variables so Claude Code talks to Z.AI:

• ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic
• ANTHROPIC_AUTH_TOKEN=$ZAI_API_KEY
• claude-glm and claude-glm-53 map Opus, Sonnet, and Haiku to glm-5.3
• claude-glm-53-flash maps all three to glm-5.3-flash

You can override the defaults by setting any of these:

• GLM_OPUS_MODEL
• GLM_SONNET_MODEL
• GLM_HAIKU_MODEL

For LiteLLM, set LITELLM_BASE_URL to the proxy root URL, LITELLM_API_KEY to its proxy key, and LITELLM_MODEL to a model_name exposed by the proxy. You can put these in the current directory's .env, in the installed switcher's .env, or export them in your shell. The LiteLLM launcher sets ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN from those values. The proxy must expose its Anthropic-compatible endpoint.

export LITELLM_BASE_URL="http://localhost:4000"
export LITELLM_API_KEY="YOUR_PROXY_KEY"
export LITELLM_MODEL="YOUR_PROXY_MODEL_NAME"
claude-litellm

claude-litellm maps Opus, Sonnet, and Haiku to LITELLM_MODEL. You can override individual slots with LITELLM_OPUS_MODEL, LITELLM_SONNET_MODEL, and LITELLM_HAIKU_MODEL.

Claude Code may warn that a LiteLLM model alias is missing from its model catalog and assume a 200k context window. If the model behind that alias and your proxy both support a larger window, set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the real token count in the same shell or .env. For example, Ollama lists deepseek-v4.1-flash:cloud with a 1M window, so an alias routed to that model can use CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000. This changes Claude Code's compaction limit without changing the model name sent to LiteLLM. A startup notice about an unrecognized model can still appear; the declared window is then used for compaction. Avoid CLAUDE_CODE_DISABLE_UNKNOWN_MODEL_WINDOW_ENFORCEMENT=1 unless your gateway passes through a context-limit error Claude Code can recognize, because that setting delays compaction until an API rejection. See Claude Code's context-window guidance and Ollama's model page.

🧰 Troubleshooting

• If claude-glm says the key is missing, make sure ZAI_API_KEY is exported in the shell you are using.
• If claude-litellm says its URL, key, or model is missing, set LITELLM_BASE_URL, LITELLM_API_KEY, and LITELLM_MODEL before launching it.
• If you skipped the key prompt during install, add export ZAI_API_KEY="YOUR_KEY_HERE" to your shell profile and reload the shell.
• If claude is not found, install Claude Code and make sure it is on PATH.

🧹 Uninstall

./uninstall.sh

📝 License

MIT. See LICENSE.