ModelHub

Self-hosted AI gateway for Claude Code & Codex — route to DeepSeek, Qwen, Kimi, GLM, Gemini, Groq with your own API key. Up to 97% cheaper than Anthropic. 7-day free trial.

ModelHub is a lightweight, self-hosted proxy that lets Claude Code and OpenAI Codex talk to any OpenAI-compatible model provider — especially low-cost models like DeepSeek, Qwen, Kimi, and Zhipu GLM. You bring your own API key (BYOK); ModelHub translates the Anthropic Messages API ↔ OpenAI Chat Completions format and handles tool use end-to-end.

🔑 Bring your own key (BYOK). ModelHub is a router — it does not sell or bundle API credits. You pay providers (DeepSeek, OpenRouter, etc.) directly with your own key. Many providers offer their own free tiers.

✨ Why ModelHub?

• Slash your API bill — run Claude Code on DeepSeek / Qwen / Gemini for up to 97% less than Anthropic's API. You pay the provider directly via your own key.
• Claude Code & Codex ready — full Anthropic ↔ OpenAI protocol conversion with complete tool_use support (file read/write, shell commands, sub-agents all work).
• Multi-model router — 8 providers, 21+ models; switch at runtime in the Web UI without restarting Claude Code.
• Zero dependencies — a single Node.js binary. No Docker, no Python, no npm install required to run.
• Web management UI — set keys, switch models, view logs in your browser at http://127.0.0.1:4000.
• Privacy-first — runs 100% on your machine; your keys never leave localhost.
• Windows & macOS — first-class Windows support; paths/encoding edge cases handled.

📦 Install

Option A — Download the release (recommended, includes the activate tool):

1. Go to Releases and grab the latest build for your OS.
2. Unzip and run modelhub (or modelhub.exe). A 7-day free trial starts automatically — no signup needed.

Option B — npm:

npm install -g modelhub-cli
modelhub start

(Requires Node ≥ 14.)

Option C — Claude Code / CodeBuddy plugin marketplace:

/plugin marketplace add wind33441998/modelhub
/plugin install anymodel-for-claude-code@anymodel-marketplace

🚀 Quick start

# 1. Start the proxy (Web UI at http://127.0.0.1:4000)
modelhub start

# 2. Add your provider API key (BYOK) — e.g. DeepSeek
modelhub keys set deepseek sk-xxxxxxxx

# 3. Open the Web UI to pick a model
modelhub ui

Then point Claude Code at the local proxy.

Environment variables

export ANTHROPIC_API_URL=http://127.0.0.1:4000
export ANTHROPIC_AUTH_TOKEN=sk-local-proxy
export ANTHROPIC_MODEL=default

Or ~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_API_URL": "http://127.0.0.1:4000",
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:4000",
    "ANTHROPIC_AUTH_TOKEN": "sk-local-proxy",
    "ANTHROPIC_MODEL": "default",
    "ANTHROPIC_SMALL_FAST_MODEL": "default",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "default",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "default",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "default",
    "CLAUDE_CODE_SUBAGENT_MODEL": "default"
  }
}

💡 Use 127.0.0.1 (not localhost) to avoid IPv6 ::1 resolution issues. Set ANTHROPIC_MODEL=default to hot-swap models from the Web UI without restarting. Fully quit and reopen Claude Code after saving.

ModelHub also targets OpenAI Codex — point Codex at the same proxy as an OpenAI-compatible backend.

🔌 Supported providers

Provider	Region	Sample models	Env var
🐋 DeepSeek	China	deepseek-chat, deepseek-reasoner	DEEPSEEK_KEY
🌊 SiliconFlow	China	Qwen, DeepSeek, GLM hosted	SILICONFLOW_KEY
🧠 Zhipu GLM	China	glm-4-plus, glm-4-air	ZHIPU_KEY
🌙 Kimi (Moonshot)	China	moonshot-v1-8k	MOONSHOT_KEY
🐱 Qwen (Alibaba)	China	qwen-max, qwen-plus, qwen-turbo	QWEN_KEY
🌐 OpenRouter	Global	200+ models	OPENROUTER_KEY
✨ Google Gemini	Global	gemini-2.5-pro, 2.5-flash	GEMINI_KEY
⚡ Groq	Global	Llama 3, DeepSeek-R1	GROQ_KEY

Bring your own key for any of these — ModelHub routes requests and converts protocols. The default / auto aliases resolve to whatever model you've selected in the Web UI; the built-in echo alias is a key-free loopback for testing the proxy chain.

💳 Pricing & License

ModelHub is free to try for 7 days (one device, full features). After the trial, choose a license:

Plan	Price	Duration	Devices
Trial	Free	7 days	1
Monthly	$3	30 days	3
Yearly	$29	365 days	5
Lifetime	$69	Forever	10

Get a license from the official site or Gumroad (links in Releases), then activate:

modelhub activate <YOUR_LICENSE_KEY>

Licenses are BYOK — you always use your own provider API keys. ModelHub is the gateway, not a credit reseller.

🛠 CLI reference

Command	Description
modelhub start [-d] [-p PORT]	Start proxy (-d daemon, -p port)
modelhub stop	Stop background proxy
modelhub status	Show status + current model
modelhub models	List available models
modelhub switch <model>	Switch active model
modelhub keys [list]	List configured keys
modelhub keys set <provider> <key>	Set a provider key
modelhub keys del <provider>	Delete a provider key
modelhub ui	Open Web UI in browser
modelhub doctor	Environment self-check
modelhub selftest	Run proxy chain self-test (no key needed)
modelhub activate <key>	Activate a license

🏗 How it works

Claude Code / Codex
   │  Anthropic Messages API
   ▼
ModelHub proxy  (http://127.0.0.1:4000)
   │  OpenAI Chat Completions
   ▼
┌──────────────┴──────────────┐
Domestic providers        Global providers
(DeepSeek, Qwen,          (OpenRouter, Gemini,
 Kimi, GLM — no VPN)       Groq)

ModelHub converts Anthropic Messages API requests into OpenAI Chat Completions, forwards them upstream, then streams the response back as Anthropic SSE — with bidirectional tool_use conversion so Claude Code's core capabilities (shell, file I/O, sub-agents) keep working.

Management API

Method	Path	Description
GET	/ , /ui	Web UI
GET	/api/models	Current model + providers
GET	/api/status	Health check
GET	/api/selftest	Key-free loopback test
POST	/api/switch	Switch model
POST	/api/keys	Save / clear keys
POST	/v1/messages	Anthropic proxy entry (used by Claude Code)

🧑‍💻 Local development

git clone https://github.com/wind33441998/modelhub.git
cd modelhub
node bin/modelhub.js start

Repackage the Skill: python build_skill.py.

📂 Repo structure

modelhub/
├── bin/modelhub.js          # CLI entry
├── lib/cli.js               # CLI dispatch (11 commands)
├── lib/proxy.js             # Proxy core (protocol + UI API)
├── assets/config.json       # Provider config (8 providers, 21+ models)
├── assets/ui.html           # Web UI
├── package.json             # npm package
├── plugins/                 # Skill form factor
└── README.md

📄 License

MIT — see LICENSE.