Claude model proxy

中文文档

Claude Desktop can point its gateway at this proxy while requests are routed by model name to DeepSeek, Moonshot/Kimi, GLM, Xiaomi MiMo, OpenAI, Gemini, Qwen, or Anthropic upstreams. The default mappings are:

Claude model	Upstream provider	Upstream model
claude-haiku-4-5	Anthropic	claude-haiku-4-5
claude-sonnet-4-6	Anthropic	claude-sonnet-4-6
claude-opus-4-7	Anthropic	claude-opus-4-7
claude-deepseek-v4-flash	DeepSeek	deepseek-v4-flash
claude-deepseek-v4-pro	DeepSeek	deepseek-v4-pro
claude-kimi-k2.6	Moonshot/Kimi	kimi-k2.6
claude-glm-4.5-air	GLM	glm-4.5-air
claude-glm-4.7	GLM	glm-4.7
claude-glm-5.1	GLM	glm-5.1
claude-mimo-v2-flash	Xiaomi MiMo	mimo-v2-flash
claude-mimo-v2-pro	Xiaomi MiMo	mimo-v2-pro
claude-mimo-v2.5-pro	Xiaomi MiMo	mimo-v2.5-pro
claude-qwen-flash	Qwen	qwen-flash
claude-qwen-plus	Qwen	qwen-plus
claude-qwen-max	Qwen	qwen-max
claude-gpt-5.4-mini	OpenAI	gpt-5.4-mini
claude-gpt-5.4	OpenAI	gpt-5.4
claude-gpt-5.5	OpenAI	gpt-5.5
claude-gemini-3.1-flash-lite-preview	Gemini	gemini-3.1-flash-lite-preview
claude-gemini-3-flash-preview	Gemini	gemini-3-flash-preview
claude-gemini-3.1-pro-preview	Gemini	gemini-3.1-pro-preview

The Claude model value is both the request model and the default response alias. When multiple request aliases share one upstream model, responses are rewritten back to the request alias used for that call. Provider aliases use claude- plus the actual upstream model name. The original Claude model names claude-haiku-4-5, claude-sonnet-4-6, and claude-opus-4-7 are sent to the Anthropic provider directly. OpenAI, Gemini, and Qwen models are not supported by the Anthropic provider.

Requirements

• Node.js 18 or newer
• Claude Desktop with Gateway / third-party inference configuration
• At least one provider API key for the models you plan to use

Run

cp .env.example .env
# Edit .env and fill in the provider keys you need.
set -a
. ./.env
set +a
npm start

If npm start fails with env: node: No such file or directory, use the included launcher instead. It uses the default node on PATH; if Node.js 18+ is not available on macOS and Homebrew is installed, it attempts brew install node automatically.

export DEEPSEEK_API_KEY="sk-..."
export MOONSHOT_API_KEY="sk-..."
export GLM_API_KEY="sk-..."
export XIAOMI_API_KEY="sk-..."
export DASHSCOPE_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."
export GEMINI_API_KEY="..."
./start.sh

start.sh starts the proxy in the background by default and writes claude-model-proxy.pid plus claude-model-proxy.log. Common commands:

./start.sh status
./start.sh stop
./start.sh restart
./start.sh foreground

Set CLAUDE_MODEL_PROXY_AUTO_INSTALL_NODE=0 before running ./start.sh to disable automatic Homebrew install attempts.

The proxy listens locally on the same URL you should configure as the gateway base URL:

http://127.0.0.1:8787

Configuration

Environment variables:

• BASE_URL: gateway-facing base URL. Default: http://127.0.0.1:8787.
• PORT: local listen port. Default: 8787.
• ADVANCED_ENV: JSON object used by the extension installer for optional provider keys and advanced overrides.
• DEEPSEEK_BASE_URL: DeepSeek-compatible API base URL. Default: https://api.deepseek.com/anthropic.
• DEEPSEEK_API_KEY: DeepSeek API key.
• MOONSHOT_BASE_URL: Moonshot/Kimi-compatible API base URL. Default: https://api.moonshot.cn/anthropic.
• MOONSHOT_API_KEY: Moonshot/Kimi API key.
• GLM_BASE_URL: Z.AI/GLM-compatible API base URL. Default: https://api.z.ai/api/anthropic.
• GLM_API_KEY: Z.AI/GLM API key. ZAI_API_KEY and ZHIPU_API_KEY are also accepted aliases.
• XIAOMI_BASE_URL: Xiaomi MiMo-compatible API base URL. Default: https://api.xiaomimimo.com/anthropic.
• XIAOMI_API_KEY: Xiaomi MiMo API key. MIMO_API_KEY is also accepted.
• ANTHROPIC_BASE_URL: Anthropic Messages API base URL. Default: https://api.anthropic.com.
• OPENAI_BASE_URL: OpenAI Chat Completions API base URL. Default: https://api.openai.com/v1.
• GEMINI_BASE_URL: Gemini OpenAI-compatible API base URL. Default: https://generativelanguage.googleapis.com/v1beta/openai.
• QWEN_BASE_URL: Qwen/DashScope OpenAI-compatible API base URL. Default: https://dashscope.aliyuncs.com/compatible-mode/v1.
• ANTHROPIC_API_KEY: Anthropic API key. Sent as x-api-key.
• OPENAI_API_KEY: OpenAI API key.
• GEMINI_API_KEY: Gemini API key. GOOGLE_API_KEY is also accepted.
• QWEN_API_KEY: Qwen/DashScope API key (or DASHSCOPE_API_KEY).
• MODEL_MAP: request model name -> upstream model name. JSON object or from=to,from2=to2.
• MODEL_ALIASES: upstream model name -> Claude response alias. JSON object or from=to,from2=to2.
• MODEL_ROUTES: upstream model name -> provider name. Provider names are deepseek, moonshot, glm, xiaomi, openai, gemini, qwen, and anthropic.
• REWRITE_RESPONSES: set to false to stop rewriting response model names.

Default mapping values:

MODEL_MAP='{"claude-deepseek-v4-flash":"deepseek-v4-flash","claude-deepseek-v4-pro":"deepseek-v4-pro","claude-kimi-k2.6":"kimi-k2.6","claude-glm-4.5-air":"glm-4.5-air","claude-glm-4.6":"glm-4.6","claude-glm-4.7":"glm-4.7","claude-glm-5":"glm-5","claude-glm-5.1":"glm-5.1","claude-mimo-v2-flash":"mimo-v2-flash","claude-mimo-v2-pro":"mimo-v2-pro","claude-mimo-v2.5-pro":"mimo-v2.5-pro","claude-mimo-v2-omni":"mimo-v2-omni","claude-gpt-5.5":"gpt-5.5","claude-gpt-5.4":"gpt-5.4","claude-gpt-5.4-mini":"gpt-5.4-mini","claude-gemini-3.1-pro-preview":"gemini-3.1-pro-preview","claude-gemini-3-flash-preview":"gemini-3-flash-preview","claude-gemini-2.5-pro":"gemini-2.5-pro","claude-gemini-2.5-flash":"gemini-2.5-flash","claude-gemini-3.1-flash-lite-preview":"gemini-3.1-flash-lite-preview","claude-gemini-2.0-flash":"gemini-2.0-flash","claude-qwen-flash":"qwen-flash","claude-qwen-plus":"qwen-plus","claude-qwen-max":"qwen-max","claude-haiku-4-5":"claude-haiku-4-5","claude-sonnet-4-6":"claude-sonnet-4-6","claude-opus-4-7":"claude-opus-4-7","claude-sonnet-4-5":"claude-sonnet-4-5","claude-opus-4-1":"claude-opus-4-1"}'
MODEL_ALIASES='{"deepseek-v4-flash":"claude-deepseek-v4-flash","deepseek-v4-pro":"claude-deepseek-v4-pro","kimi-k2.6":"claude-kimi-k2.6","glm-4.5-air":"claude-glm-4.5-air","glm-4.6":"claude-glm-4.6","glm-4.7":"claude-glm-4.7","glm-5":"claude-glm-5","glm-5.1":"claude-glm-5.1","mimo-v2-flash":"claude-mimo-v2-flash","mimo-v2-pro":"claude-mimo-v2-pro","mimo-v2.5-pro":"claude-mimo-v2.5-pro","mimo-v2-omni":"claude-mimo-v2-omni","gpt-5.5":"claude-gpt-5.5","gpt-5.4":"claude-gpt-5.4","gpt-5.4-mini":"claude-gpt-5.4-mini","gemini-3.1-flash-lite-preview":"claude-gemini-3.1-flash-lite-preview","gemini-3-flash-preview":"claude-gemini-3-flash-preview","gemini-3.1-pro-preview":"claude-gemini-3.1-pro-preview","gemini-2.5-pro":"claude-gemini-2.5-pro","gemini-2.5-flash":"claude-gemini-2.5-flash","gemini-2.0-flash":"claude-gemini-2.0-flash","qwen-flash":"claude-qwen-flash","qwen-plus":"claude-qwen-plus","qwen-max":"claude-qwen-max","claude-haiku-4-5":"claude-haiku-4-5","claude-sonnet-4-6":"claude-sonnet-4-6","claude-opus-4-7":"claude-opus-4-7","claude-sonnet-4-5":"claude-sonnet-4-5","claude-opus-4-1":"claude-opus-4-1"}'
MODEL_ROUTES='{"deepseek-v4-flash":"deepseek","deepseek-v4-pro":"deepseek","kimi-k2.6":"moonshot","glm-4.5-air":"glm","glm-4.6":"glm","glm-4.7":"glm","glm-5":"glm","glm-5.1":"glm","mimo-v2-flash":"xiaomi","mimo-v2-pro":"xiaomi","mimo-v2.5-pro":"xiaomi","mimo-v2-omni":"xiaomi","gpt-5.5":"openai","gpt-5.4":"openai","gpt-5.4-mini":"openai","gemini-3.1-pro-preview":"gemini","gemini-3-flash-preview":"gemini","gemini-2.5-pro":"gemini","gemini-2.5-flash":"gemini","gemini-3.1-flash-lite-preview":"gemini","gemini-2.0-flash":"gemini","qwen-flash":"qwen","qwen-plus":"qwen","qwen-max":"qwen","claude-haiku-4-5":"anthropic","claude-sonnet-4-6":"anthropic","claude-opus-4-7":"anthropic","claude-sonnet-4-5":"anthropic","claude-opus-4-1":"anthropic"}'

Health check

curl http://127.0.0.1:8787/healthz

Claude Code

Claude Code can use this proxy through its Anthropic-compatible environment variables. Start the proxy first, then source the Claude Code client environment in the terminal where you run claude:

cp .env.claude-code.example .env.claude-code
# Edit .env.claude-code if you want different model aliases.
set -a
. ./.env.claude-code
set +a
claude

The default Claude Code example maps its aliases to these proxy models:

ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-deepseek-v4-flash
ANTHROPIC_DEFAULT_SONNET_MODEL=claude-deepseek-v4-pro
ANTHROPIC_DEFAULT_OPUS_MODEL=claude-kimi-k2.6
CLAUDE_CODE_SUBAGENT_MODEL=claude-deepseek-v4-flash
ANTHROPIC_MODEL=sonnet

You can also start Claude Code with a specific proxy model directly:

ANTHROPIC_BASE_URL=http://127.0.0.1:8787 \
ANTHROPIC_API_KEY=dummy-claude-model-proxy \
claude --model claude-deepseek-v4-pro

ANTHROPIC_API_KEY is only a non-empty client-side placeholder for this proxy. Provider API keys still come from the proxy's .env, MCPB install settings, or LaunchAgent environment file.

Claude Code works best with Anthropic Messages-compatible upstreams such as DeepSeek, Moonshot/Kimi, GLM, Xiaomi MiMo, and Anthropic because tool-use payloads are passed through as-is. OpenAI, Gemini, and Qwen routes use a basic Chat Completions adapter for text/image and streaming responses; they are not a full Claude Code tool-use compatibility layer.

Claude Desktop extension

Build the installable MCPB extension:

npm run build:mcpb

The output is:

dist/claude-model-proxy-0.1.0.mcpb

Install it in Claude Desktop from Settings -> Extensions / Connectors -> Advanced settings -> Install Extension. During first installation, fill in the gateway URL, local port, DeepSeek credentials, and Moonshot/Kimi credentials. Optional providers and mapping overrides can be supplied through the advanced JSON field.

After the extension is installed, open Claude Desktop settings and configure third-party inference as shown below. If the third-party inference or extension controls are hidden, enable Developer Mode first.

Use these values:

• Provider: Gateway
• Gateway base URL: http://127.0.0.1:8787
• Gateway API key: any non-empty placeholder, for example dummy-claude-model-proxy
• Gateway auth scheme: bearer
• Model list: add the Claude-style request model names you want to expose, such as claude-deepseek-v4-flash, claude-deepseek-v4-pro, and claude-kimi-k2.6, claude-qwen-flash, claude-qwen-plus, or claude-qwen-max

Provider API keys are configured in the extension installer or environment variables, not in the Gateway API key field.

The extension exposes a model_proxy_status tool so you can inspect local proxy status, providers, and model mappings from Claude.

In Claude Desktop settings, this appears under Tool permissions as Other tools -> Model proxy status. If its permission is Needs approval, Claude will ask before each tool call. The tool is annotated as read-only and non-destructive, so new installs on clients that honor MCP annotations should be able to default it to no approval; existing user permission overrides may need to be changed once in settings. In a Claude chat, ask:

Use the Model proxy status tool to check whether Claude Model Proxy is running.

or:

Call model_proxy_status and tell me whether the proxy is listening and whether
API keys are configured.

The returned JSON includes listening, error, external, localUrl, provider hasApiKey flags, and the active model mappings. You can also inspect similar status from a terminal:

curl http://127.0.0.1:8787/healthz

If a brand-new chat says model_proxy_status is unavailable but the same request works immediately on retry, Claude started the model response before it had refreshed the MCPB tool list. The extension declares a fixed static tool list in the bundle metadata and marks model_proxy_status as read-only so clients can discover and approve it early. If the first already-started response was built without tools, retry the request once or open the chat after Claude Desktop has finished loading the extension. Keeping the proxy itself running as a LaunchAgent also avoids the separate gateway-listening race described below.

Avoid the startup warning

Claude Desktop can probe the gateway before the MCPB extension process has finished starting. When that happens, Claude shows "Can't reach 127.0.0.1:8787" until you click "Check again".

For a clean startup, run the proxy as a macOS LaunchAgent so it is already listening before Claude starts:

npm run launch-agent:install

The installer creates ~/.claude-model-proxy.env. Put the same provider keys in that file, then restart the agent:

launchctl kickstart -k gui/$(id -u)/local.claude-model-proxy
curl http://127.0.0.1:8787/healthz

The MCPB extension can still stay installed. If it sees the LaunchAgent already owning port 8787, its status reports the proxy as externally running instead of treating the port conflict as a failure.

To remove the LaunchAgent:

npm run launch-agent:uninstall

Extension install UI

The MCPB installer shows only the gateway/proxy basics plus DeepSeek and Moonshot/Kimi credentials by default. Less common provider credentials and mapping overrides stay available through one optional advanced JSON field:

{
  "GLM_API_KEY": "...",
  "XIAOMI_API_KEY": "...",
  "ANTHROPIC_API_KEY": "sk-ant-...",
  "OPENAI_API_KEY": "sk-...",
  "GEMINI_API_KEY": "...",
  "QWEN_API_KEY": "sk-..."
}

The advanced field accepts any environment variable listed below, including MODEL_MAP, MODEL_ALIASES, MODEL_ROUTES, and REWRITE_RESPONSES.

Project layout

.
├── manifest.json              # MCPB extension manifest
├── proxy.mjs                  # HTTP gateway proxy and provider adapters
├── server/index.mjs           # MCP stdio server that starts the proxy
├── scripts/                   # Build, launchd, and Node helper scripts
├── srcs/                      # README screenshots and images
├── test/proxy.test.mjs        # Node test suite
├── start.sh                   # Standalone launcher
├── .env.claude-code.example   # Claude Code client configuration template
└── .env.example               # Safe local configuration template

Generated files under dist/, local .env files, logs, editor files, and dependencies are ignored by .gitignore.

Notes

DeepSeek, Moonshot/Kimi, GLM, Xiaomi MiMo, and Anthropic are treated as Anthropic Messages-compatible upstreams. OpenAI, Gemini, and Qwen are adapted through their OpenAI-compatible Chat Completions endpoints. The adapter covers normal text and image message content plus streaming text deltas; Anthropic tool-use blocks, audio, and provider-specific advanced options are intentionally left as upstream-specific behavior.