ccmm — Claude Code Model Manager

The fastest way to switch Claude Code's model. Define your own model plans — map each thinking depth (auto / high / medium / low / subagent) to whatever model you want, across any provider. Then !ccmm use deepseek switches all 5 slots at once. Live, mid-session, no restart.

Bonus: real-time cost metering, cache-hit tracking, budget alerts — all in the status line.

< English | 简体中文 >

┌──────────────┐   ANTHROPIC_BASE_URL    ┌──────────────┐        ┌─────────────────────┐
│  Claude Code │ ───────────────────────▶│  ccmm proxy  │───────▶│ Anthropic / relay / │
│    (CLI)     │◀─────────────────────── │  (localhost) │◀───────│ OpenAI-compat / local│
└──────────────┘                         └──────┬───────┘        └─────────────────────┘
        ▲  status line                          │ rewrites `model`, meters usage, prices it
        └──────────── reads live metrics ───────┘  ~/.ccmm/ (config · route · metrics)

---

❓ Why ccmm?

Claude Code picks models automatically based on thinking depth, but it doesn't let you choose which models those are — and it sure doesn't let you switch all five at once.

ccmm flips that: you define a 方案 (plan) — a named set of five model slots mapping each thinking depth (auto / high / medium / low / subagent) to whatever model you want, on whatever provider you trust. Then one command swaps the entire plan. Live. Mid-session. No restart.

v1 is Anthropic Messages API only. Your provider must speak the Anthropic wire protocol. OpenAI-compatible translation is on the roadmap.

Beyond switching:

• Your cost, not Anthropic's — if you route through a different provider, Claude's built-in cost number is wrong. ccmm meters at the proxy on the actual forwarded model with your own price table.
• Cache-hit visibility — ccmm tracks cache_read_input_tokens and cache_creation_input_tokens from SSE events and shows the real cache-hit percentage in the status line. You can see whether your provider is giving you the caching you're paying for.
• No workflow interruption — !ccmm use my-model in the Claude Code prompt box. Hot-reloaded. Instant.

✨ Features

🔀 Live switching	!ccmm use deepseek inside Claude Code — next message uses the new provider, no restart.
📊 Real-time status line	Active model · tokens in/out · cache-hit % · today cost · budget remaining.
🌐 Any provider	Anthropic, DeepSeek, OpenRouter, Vercel AI, Moonshot, or any Anthropic-compatible endpoint.
🗂 Named provider plans	One ccmm use <name> switches an entire 5-slot model mapping at once.
💵 Your prices	Per-model price tables so cost is accurate even behind a relay.
🚦 Budgets & alerts	Daily spend caps visible in the status line.
🎛 Interactive config	ccmm config — menu-driven editor for all settings.
🌍 Bilingual	中文 / English — choose on ccmm setup, switch anytime in ccmm config.
🚀 Auto-start	Optional: auto-launch the proxy daemon on system login.
🔄 Auto-update	ccmm update checks npm for new versions; auto-notifies on start/config/setup.
🧩 Plugin	Auto-registers status line and MCP tools on install.

📋 Prerequisites

• Node.js >= 18.0.0
• Claude Code installed (@anthropic-ai/claude-code)

📦 Install

npm install -g @pgwtf/ccmm
ccmm setup         # interactive wizard: language → providers → 5-slot model map → done

ccmm setup handles everything (init, provider config, settings sync). It's idempotent and backs up every file it touches.

🚀 Quickstart

ccmm setup                # first-time wizard (skip if already configured)
ccmm config               # interactive editor — tweak providers, prices, budget anytime
ccmm start                # start the proxy daemon
ccmm use deepseek         # switch to DeepSeek (`!ccmm use deepseek` inside Claude)
ccmm stats today          # what did I spend?
ccmm update               # check & install updates

📟 Commands

Command
ccmm setup	Interactive wizard — add providers, 5-slot model mapping, budget, auto-start
ccmm config	Interactive editor — browse/modify all settings
ccmm start / stop / restart / logs	Proxy daemon lifecycle
ccmm use <name>	Switch active provider/方案 (live, hot-reloaded)
ccmm current	Show active provider and model
ccmm models	List all providers and priced models
ccmm provider add|rm|list	Manage providers via CLI
ccmm price set|rm|list	Manage per-model prices
ccmm stats [today|session|week|all]	Usage & cost report
ccmm statusline	(internal) renders the status line
ccmm doctor	Diagnose setup
ccmm update	Check for updates and install the latest version
ccmm init	Quick non-interactive init (prefer ccmm setup)

⚙️ Configuration

~/.ccmm/config.json:

{
  "proxy":   { "host": "127.0.0.1", "port": 8787 },
  "defaultProvider": "deepseek",
  "language": "zh-CN",
  "providers": {
    "deepseek": {
      "baseUrl": "https://api.deepseek.com/anthropic",
      "apiKeyEnv": "DEEPSEEK_API_KEY",
      "wire": "anthropic",
      "modelMap": {
        "ANTHROPIC_MODEL": "deepseek-v4-pro",
        "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-v4-pro",
        "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4-flash",
        "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-v4-flash",
        "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek-v4-flash"
      }
    }
  },
  "pricesUSD": {
    "deepseek-v4-pro": { "input": 2.0, "output": 8.0, "cacheRead": 0.2, "cacheWrite": 3.0 }
  },
  "pricesCNY": {},
  "budget": { "daily": 20, "alert": true }
}

• Provider/方案 — a named config (endpoint + API key + 5-slot modelMap). One ccmm use <name> switches the entire plan.
• modelMap — maps Claude Code's 5 thinking-depth placeholders to real upstream model IDs.
• Prices — per 1M tokens, stored per currency (pricesUSD / pricesCNY). Cost is computed from the forwarded model, not the requested one.
• currency — "USD" or "CNY". Determines which price table and symbol ($/¥) is used.
• language — "zh-CN" or "en".

🔧 How it works

1. ccmm setup writes ANTHROPIC_BASE_URL=http://127.0.0.1:8787 into Claude Code settings.
2. All Claude Code API requests go through the ccmm proxy.
3. The proxy checks the active provider's modelMap → rewrites body.model → forwards.
4. Response streams back transparently; usage is captured from SSE events for metering.
5. Prompt caching is preserved — only body.model and auth headers are touched.

📊 Status line

🧠 deepseek-v4-pro · ▲12.4k ▼3.1k · cache 87% · $0.42 today · $19.58 left

Auto-degrades to parsing Claude Code's transcript JSONL when the proxy isn't running.

🗺️ Roadmap

Interactive TUI dashboard, model recommendations, Bedrock/Vertex providers, OpenAI-compatible wire translation.

🙏 Credits

• tweakcc · claude-code-router · LiteLLM
• How I built a hot-swappable backend proxy for Claude Code
• Anthropic's LLM gateway docs

⚖️ Disclaimer

ccmm is an independent project — not affiliated with Anthropic. API keys are stored locally in ~/.ccmm/ and sent only to the providers you specify.

📄 License

MIT (LICENSE).