Claude Model Switcher

If this tool helps you, consider giving it a ⭐ on GitHub — it means a lot!

中文文档

A lightweight CLI that lets you switch AI model providers in Claude Code with a simple @provider syntax.

Each terminal session runs independently — open multiple terminals and use different providers or models simultaneously, with zero conflict.

Quick Start

ccs @zhipu              # Use Zhipu GLM
ccs @zhipu:glm-4.6      # Specify model
ccs                      # Default provider

Installation

npm install -g @vinoorg/claude-model-switcher

Requirements: Bun >= 1.0, Claude Code CLI installed

Configuration

1. Create provider config

mkdir -p ~/.config/claude-model-switcher
cat > ~/.config/claude-model-switcher/providers.json << 'EOF'
[
  {
    "name": "zhipu",
    "base_url": "https://open.bigmodel.cn/api/anthropic",
    "api_key_env": "ZHIPU_API_KEY",
    "default_model": "glm-4.6",
    "default_small_model": "glm-4.5-air"
  },
  {
    "name": "lp",
    "base_url": "http://your-company.com/api/anthropic",
    "api_key_env": "LP_API_KEY"
  }
]
EOF

2. Set API keys

# ~/.zshrc or ~/.bashrc
export ZHIPU_API_KEY="your-key-here"
export LP_API_KEY="your-key-here"

3. Apply changes

source ~/.zshrc  # or source ~/.bashrc

Note: ccs transparently passes all arguments to the native claude CLI. Your existing Claude Code commands and flags work exactly as before — ccs only sets model-related environment variables.

Usage

ccs                          # Default provider
ccs @zhipu                   # Specific provider, default model
ccs @zhipu:glm-4.6           # Provider + model
ccs @list                    # List all providers & models
ccs @init                    # Init or validate config
ccs @update                  # Update to latest version
ccs @config                  # Open config file in editor
ccs @trace                   # Open traces directory in editor
ccs @trace latest            # Open the most recent trace session
ccs @trace <session-id>      # Open a specific trace session
ccs @trace clean             # Keep the latest 50 sessions, remove older ones
ccs @trace clean --all       # Remove all trace sessions
ccs @trace clean --keep N    # Keep the latest N sessions
ccs @help                    # Show help

# Works with all native Claude Code flags
ccs @zhipu -r sessionID
ccs @lp -p "Describe this project"

Status Bar Companion

Display the current provider, model, and context usage in your Claude Code status bar with ccs-statusline:

[v2.1.119]  ➜ my-project (main) @zhipu:glm-4.6  Ctx:45.2k/200k (22.6%)

Troubleshooting

Still connecting to an old API endpoint after using ccs

If you previously used tools like cc-switch that set ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL, or ANTHROPIC_AUTH_TOKEN in your ~/.zshrc / ~/.bashrc, remove or comment out these variables. Otherwise they will be inherited by the child process and override ccs configuration.

Config Reference

| Field | Required | Description | | ------------------- | -------- | ---------------------------------------------------------------------------------------------- | | name | Yes | Provider identifier for @name. Only [a-zA-Z0-9_-]. Reserved: list, help, init, update, config | | base_url | Yes | Anthropic-compatible API endpoint | | api_key_env | Yes | Environment variable name holding the API key | | default_model | No | Model used when not specified | | default_small_model | No | Lightweight model (falls back to default_model) | | models | No | Available model list (string[]). Defaults to showing default_model + default_small_model | | trace | No | Set to true to record this provider's API traffic to ~/.config/claude-model-switcher/traces/ for debugging |

Rules:

• Without @provider, the first entry is used as default
• If default_model is not set, model must be specified explicitly (ccs @lp:some-model)
• default_small_model only affects ANTHROPIC_SMALL_FAST_MODEL and ANTHROPIC_DEFAULT_HAIKU_MODEL

Adding a new provider — just append to the config array and export the API key:

{
  "name": "kimi",
  "base_url": "https://api.kimi.com/anthropic",
  "api_key_env": "KIMI_API_KEY",
  "default_model": "kimi-latest"
}

No code changes needed.

Tracing API Traffic

Set "trace": true on a provider to record every Claude Code API request/response to that provider for debugging. Traces are written to ~/.config/claude-model-switcher/traces/ as JSONL files named <yyyymmddHHMMSS>-<provider>-<8hex>.jsonl (e.g. 20260701143022-zhipu-a1b2c3d4.jsonl). Sensitive request/response headers are redacted.

{
  "name": "zhipu",
  "base_url": "https://open.bigmodel.cn/api/anthropic",
  "api_key_env": "ZHIPU_API_KEY",
  "default_model": "glm-4.6",
  "trace": true
}

Inspect recorded sessions with ccs @trace (see Usage). By default the 50 most recent sessions are kept; older ones are pruned automatically when a new session is recorded. Use ccs @trace clean [--all|--keep N] to prune manually.

Note: a trace record's incomplete flag does not capture every upstream interruption. Bun.fetch (WHATWG Streams) may surface a mid-stream upstream break as a clean EOF, in which case incomplete stays false; the flag only fires when the proxy's stream read is rejected (e.g. TCP reset).

License

MIT