@shuhang/ccmodel

Interactive CLI to switch Claude Code model providers and launch claude with the right environment. Run npx @shuhang/ccmodel, pick a provider, enter an API key, select a model — Claude Code starts up configured.

Quick Start

npx @shuhang/ccmodel

No global install required.

CLI Flags

ccmodel                 # start the interactive menu
ccmodel --version       # print version and exit (alias: -v)
ccmodel --help          # print usage and exit   (alias: -h)

Any extra arguments after ccmodel are forwarded to claude once you launch, e.g. ccmodel mcp.

What It Does

  Claude Code Model Launcher

? 选择操作:
❯ 选择模型并启动 Claude Code
  添加/修改 API Key
  退出

Select a provider → enter API key → pick a model → Claude Code launches with the correct environment variables. The parent process waits for claude to exit and then exits with the same code — the launcher never shares the terminal with Claude Code, so there is no UI overlap.

Supported Providers

21 provider profiles across domestic (CN) and international endpoints. Providers marked ⚠ have endpoints that are not protocol-compatible with the Anthropic API out of the box, or are not fully verified for Claude Code; the menu warns and asks for confirmation before launching them.

| Provider | Region | Base URL | Models | |----------|--------|----------|--------| | Anthropic | Global | api.anthropic.com | Claude Opus 4.7, Sonnet 4.6, Haiku 4.5 | | DeepSeek | Global | api.deepseek.com/anthropic | DeepSeek V4 Pro, V4 Flash | | GLM (智谱) | CN | open.bigmodel.cn/api/anthropic | GLM-5.1 | | GLM (智谱) | International | api.z.ai/api/anthropic | GLM-5.1 | | Qwen (通义千问) | CN | dashscope.aliyuncs.com/apps/anthropic | Qwen3 Max, Qwen3.5 Plus, Qwen Plus, Qwen3 Coder Next, Qwen3 Coder Plus | | Qwen (百炼 Coding) | CN | coding.dashscope.aliyuncs.com/apps/anthropic | Qwen3 Coder Plus | | Kimi (月之暗面) | CN | api.moonshot.cn/anthropic | Kimi K2.6, K2.5 | | Kimi (Coding Plan) | CN | api.kimi.com/coding/ | custom — prompted at launch | | MiniMax | CN | api.minimaxi.com/anthropic | MiniMax M2.7 | | MiniMax | International | api.minimax.io/anthropic | MiniMax M2.7 | | 硅基流动 SiliconFlow | CN | api.siliconflow.cn | MiniMax M2.7 Pro | | 硅基流动 SiliconFlow | International | api.siliconflow.com | MiniMax M2.7 | | 阶跃星辰 StepFun | CN | api.stepfun.com/step_plan | Step 3.5 Flash | | 阶跃星辰 StepFun | International | api.stepfun.ai/step_plan | Step 3.5 Flash | | OpenRouter | Global | openrouter.ai/api | Claude Sonnet 4.6, Opus 4.7, DeepSeek V4, Kimi K2.6 | | 豆包 Doubao (字节) ⚠ | CN | ark.cn-beijing.volces.com/api/coding | Doubao Seed 2.0 Code, Pro, Lite, Mini | | 小米 MiMo | CN | api.xiaomimimo.com/anthropic | MiMo V2 Pro | | Novita AI | Global | api.novita.ai/anthropic | GLM-5.1, Kimi K2.5, Kimi K2.6 | | OpenAI ⚠ | Global | api.openai.com/v1 | GPT-5.5 Pro, GPT-5.5, GPT-5.4 Mini | | Google Gemini ⚠ | Global | generativelanguage.googleapis.com | Gemini 3.1 Pro | | GitHub Copilot ⚠ | Global | api.githubcopilot.com | Claude Sonnet 4.6 |

How It Works

1. Reads ~/.ccmodel/config.json for saved API keys and default models.

2. Asks you for a main model (chat + Opus/Sonnet tier tasks) and a fast model (Haiku tier + subagents). Fast model defaults to the main model — press Enter to skip.

3. Runs Claude Code with an isolated config directory, ~/.ccmodel/claude-config, by setting CLAUDE_CONFIG_DIR. This prevents user-level ~/.claude/settings.json from overriding the provider, API key, or model choices made in ccmodel.

4. Builds a clean child environment: inherits your shell env but strips ANTHROPIC_*, CLAUDE_CODE_SUBAGENT_MODEL, CLAUDE_CONFIG_DIR, API_TIMEOUT_MS, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC so nothing leftover can hijack the request.

5. Sets the full set of Claude Code model-routing variables so subagent and background tasks also go to your chosen provider:

   | Variable | Value | |----------|-------| | ANTHROPIC_MODEL | main model | | ANTHROPIC_DEFAULT_OPUS_MODEL | main model | | ANTHROPIC_DEFAULT_SONNET_MODEL | main model | | ANTHROPIC_DEFAULT_HAIKU_MODEL | fast model | | CLAUDE_CODE_SUBAGENT_MODEL | fast model | | ANTHROPIC_AUTH_TOKEN | your API key | | ANTHROPIC_BASE_URL and any provider-specific env | from provider config |

   This is required for third-party providers: Claude Code internally routes different tiers of tasks (subagent spawning, background summarization, etc.) through different model aliases, and without these variables it falls back to Anthropic's default IDs (e.g. claude-haiku-4-5) that third-party endpoints will reject with "Unknown Model".

6. Resolves the claude binary path (honors CCMODEL_CLAUDE_PATH; warns if multiple claude binaries are on PATH).

7. Releases the terminal raw mode and spawns claude with inherited stdio.

8. Waits for claude to exit, then exits the launcher with the same code.

No config files in your project are touched. User-level Claude Code settings under ~/.claude/ are also left untouched.

Why two models?

Claude Code is an agent system: a single session makes many model calls. Some are the main conversation, others are background tasks like subagent spawning (Task tool), title generation, or tool-call planning. The Anthropic official docs define four routing env vars (ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL, CLAUDE_CODE_SUBAGENT_MODEL). ccmodel simplifies those into a two-tier choice:

• Main model covers ANTHROPIC_MODEL, DEFAULT_OPUS_MODEL, and DEFAULT_SONNET_MODEL.
• Fast model covers DEFAULT_HAIKU_MODEL and CLAUDE_CODE_SUBAGENT_MODEL. Pick a cheaper / faster model here when your provider offers one (e.g. deepseek-v4-flash for DeepSeek, kimi-k2-turbo-preview for Kimi). If the provider only has one model, same-as-main is fine.

Security Posture

• Masked input: API keys are read with a password prompt (displayed as *).
• Plaintext storage (0600): once validated (≥ 8 chars), keys are saved to ~/.ccmodel/config.json with mode 0600 on macOS/Linux. The content is plaintext — this is intentional (same trust level as the Claude Code CLI) but you should know.
• Shell env scrubbing: leftover ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, CLAUDE_CODE_SUBAGENT_MODEL, and similar variables in your shell cannot silently redirect your key or model to a different target; the launcher removes them before applying the provider config.
• Claude settings isolation: by default, ccmodel sets CLAUDE_CONFIG_DIR=~/.ccmodel/claude-config before launching Claude Code. This avoids reading or modifying your normal ~/.claude/settings.json, which may contain unrelated provider settings.
• Incompatible-endpoint warnings: providers whose endpoints need a protocol adapter (OpenAI, Gemini, GitHub Copilot) are marked ⚠ 不兼容 in the menu and require an extra confirm before launch.
• Binary trust: set CCMODEL_CLAUDE_PATH=/path/to/claude to pin a specific claude binary and avoid PATH shadowing. When multiple claude binaries are found, the launcher prints every candidate before using the first one.
• Corrupt config recovery: if ~/.ccmodel/config.json is unreadable or malformed, it is moved to ~/.ccmodel/config.json.corrupt-<timestamp> instead of being silently overwritten — no other providers' keys are lost.
• Ctrl+C is quiet: pressing Ctrl+C in any prompt exits with code 130 without printing a stack trace.

Configuration File

~/.ccmodel/config.json:

{
  "providers": {
    "deepseek": {
      "apiKey": "sk-..."
    },
    "glm": {
      "apiKey": "...",
      "defaultModel": "glm-5.1"
    }
  }
}

defaultModel is set interactively the first time you pick a non-default model for a multi-model provider. Next time, that model is labeled [默认] in the menu.

Environment Variables

| Variable | Purpose | |----------|---------| | CCMODEL_CLAUDE_PATH | Override claude binary resolution. Launcher validates the path is executable and warns if not. | | CCMODEL_CLAUDE_CONFIG_DIR | Override the isolated Claude Code config directory. Defaults to ~/.ccmodel/claude-config. | | CCMODEL_USE_USER_CLAUDE_CONFIG=1 | Opt out of isolation and let Claude Code use your normal ~/.claude config. Not recommended when switching providers. | | CCMODEL_DEBUG_ENV=1 | Print key environment variables passed to claude before launch. API keys/tokens are masked. |

Everything starting with ANTHROPIC_*, plus CLAUDE_CODE_SUBAGENT_MODEL, CLAUDE_CONFIG_DIR, API_TIMEOUT_MS, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, is stripped from inherited env before spawn. Only values from the chosen provider config take effect inside claude.

Requirements

• Node.js ≥ 18
• Claude Code installed: npm install -g @anthropic-ai/claude-code

Install Globally (Optional)

npm install -g @shuhang/ccmodel
ccmodel

Troubleshooting

• "Cannot find module './menu.js'" during npm run dev — use npm run dev which runs via tsx, not ts-node. ts-node under CJS cannot resolve the .js extensions used in source imports.
• claude not found — install Claude Code globally, or set CCMODEL_CLAUDE_PATH to the exact path.
• Warnings about multiple claude binaries — resolve ambiguity by setting CCMODEL_CLAUDE_PATH.
• UI looks corrupted on launch — make sure you are running in a real interactive terminal (not piped stdin) so both the menu and Claude Code can control the TTY properly.

License

MIT