cc-proxy

A Claude Code plugin with a local proxy that lets one session use GLM (Z.ai), DeepSeek, Qwen, OpenRouter, LM Studio and Claude side by side. Switch with /model, no restart. Zero runtime dependencies.

Claude Code points ANTHROPIC_BASE_URL at http://127.0.0.1:4000. The proxy routes each request by its model name, applies that backend's auth, and forwards it unchanged, so every tool, subagent and prompt cache keeps working.

What routes where

You pick	It goes to	Needs
glm-5.3[1m], any glm-*	GLM (Z.ai)	GLM_API_KEY
deepseek-v4-pro, any deepseek-*	DeepSeek	DEEPSEEK_API_KEY
qwen3.7-max, any qwen*	Qwen Token Plan	DASHSCOPE_API_KEY
deepseek/deepseek-v4-pro, any vendor/model	OpenRouter	OPENROUTER_API_KEY
lmstudio:<model-id>	your LM Studio server	LMSTUDIO_BASE_URL
opus, sonnet, any claude-*, anything else	Claude, on your own login	nothing

Every key is optional. A model served by several backends takes the native one first; a <provider>: prefix such as qwen:deepseek-v4-pro names another. The full rules, the [1m] suffix and the haiku pin are in docs/ROUTING.md.

Install

Needs Node 22 or newer on your PATH.

claude plugin marketplace add betmoar/ccp-market
claude plugin install cc-proxy@betmoar

Or straight from this repo, which carries its own marketplace manifest:

claude plugin marketplace add betmoar/cc-proxy-plugin
claude plugin install cc-proxy@cc-proxy-plugin

Setup

/cc-proxy:setup

It asks for the keys you want (all optional), then:

• writes them to ~/.env, the one file the proxy reads keys from;
• sets ANTHROPIC_BASE_URL in ~/.claude/settings.json env;
• publishes one /model picker row per routable model, each carrying its real context window (Claude Code otherwise assumes 200K for every id it does not know);
• starts the proxy, so a fresh session connects without ECONNREFUSED.

After that the SessionStart hook keeps the proxy running and replaces an older one after a plugin update. Everything it writes is described in docs/CONFIGURATION.md.

Use it

/model glm-5.3[1m]              # GLM, 1M context
/model deepseek-v4-pro          # DeepSeek
/model qwen3.7-max              # Qwen plan
/model lmstudio:openai/gpt-oss-20b
/model opus                     # back to Claude

Routing decisions land in ~/.claude/cc-proxy/cc-proxy.log, one line per request. curl http://127.0.0.1:4000/v1/models lists everything reachable; that list is a published contract other plugins consume, documented in docs/DISCOVERY.md.

Commands

Command	Does
/cc-proxy:setup	Keys, settings, picker rows, and starts the proxy
/cc-proxy:status	Proxy liveness and version, configured backends, GLM and OpenRouter quota, recent routing lines
/cc-proxy:models	Every reachable model with the backend it routes to
/cc-proxy:bench grades	Refresh model grades from benchlm.ai and OpenRouter into ~/.claude/cc-proxy/grades.json. Manual by design
/cc-proxy:bench speed	Time one turn per route; --report gives median and p95 over the series

The GLM offload subagents moved to betmoar/cc-agents-plugin.

Statusline (optional)

cc 5h:2% | glm 5h:14% | or:$$$ | ds:$$ | qw:on

Claude and GLM 5-hour usage, OpenRouter credits, DeepSeek balance, and proxy down when the port is closed. It never waits on the network. Setup offers to enable it; the format, the ! and ? marks, and what to do when a gauge sticks are in docs/STATUSLINE.md.

Docs

Read	For
docs/ROUTING.md	Which backend an id goes to, <provider>: prefixes, what the proxy changes in a request
docs/DISCOVERY.md	The /v1/models contract, ?dedup=identity, the image-generation tunnel
docs/CONFIGURATION.md	Every environment variable, auth mode for off-loopback binds, the /model picker rows
docs/STATUSLINE.md	The status bar segment
docs/TROUBLESHOOTING.md	Symptom → cause → fix
docs/OPERATIONS.md	How the plugin is installed and updated, the proxy's lifecycle and endpoints, prompt caching, files on disk
docs/ARCHITECTURE.md	Design, invariants and the reasons behind them
CONTRIBUTING.md	Local development, gates, adding a provider
docs/RELEASING.md	Merging and cutting a release
docs/MAINTAINING.md	Traps, decision procedures, the why behind every coupling
docs/BACKLOG.md	Open work, closed items with their evidence, reversed decisions
CHANGELOG.md	What changed, by version

Limitations

• macOS and Linux verified; Windows untested.
• GLM via Z.ai's Coding Plan endpoint (https://api.z.ai/api/anthropic); the Standard api/paas/v4 API is not supported.
• Relies on a few Claude Code internals ([1m] suffix, claude-haiku-* ids, the modelPicker schema) that are not public API and may drift.
• No format translation: every backend must speak the Anthropic Messages API.