Claude Alias Patch

Add custom model aliases to Claude Code. Use Gemini, GPT, or any model as a subagent — Claude sees them as selectable options alongside Sonnet, Opus, and Haiku.

How it works

Claude Code hardcodes its model aliases to sonnet, opus, and haiku. This patch extends that list by scanning for ANTHROPIC_DEFAULT_*_MODEL environment variables in your ~/.claude/settings.json.

For example, setting ANTHROPIC_DEFAULT_GEMINI_MODEL registers gemini as a model alias. Claude can then select it when spawning subagents via the Task tool, in the model picker, and in agent definitions.

If you don't set any custom env vars, the patch does nothing — Claude Code works exactly like normal.

You'll need an API proxy that accepts Anthropic-format API calls and routes them to the correct provider. I use 9router, but LiteLLM and OpenRouter work too.

Install

curl -sL https://raw.githubusercontent.com/East-rayyy/claude-alias-patch/main/scripts/install.sh | bash

Prerequisites: Node.js >= 18, npm, python3.

The installer downloads Claude Code from npm, patches it, backs up your existing claude binary, and installs a wrapper script. Works whether you originally installed Claude Code via npm or the native binary.

All scripts live in the scripts/ folder.

Configure

Add your custom models to ~/.claude/settings.json. These env vars are completely optional — only add the ones you need:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:PORT/v1",
    "ANTHROPIC_DEFAULT_GEMINI_MODEL": "google/gemini-2.5-pro",
    "ANTHROPIC_DEFAULT_GPT_MODEL": "openai/gpt-4o",
    "ANTHROPIC_DEFAULT_DEEPSEEK_MODEL": "deepseek/deepseek-r1"
  }
}

The naming convention is ANTHROPIC_DEFAULT_{ALIAS}_MODEL. The alias comes from the middle part:

• ANTHROPIC_DEFAULT_GEMINI_MODEL → alias gemini
• ANTHROPIC_DEFAULT_GPT_MODEL → alias gpt
• ANTHROPIC_DEFAULT_DEEP_SEEK_MODEL → alias deep-seek (underscores become hyphens)

Restart Claude Code after changing your env vars. Your custom aliases show up automatically in the model picker, the Task tool, and agent definitions. No need to re-patch or rerun the installer — just edit and restart.

My setup

I use 9router as my API proxy. Here's what my settings.json env looks like:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:20128/v1",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4-6",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5",
    "ANTHROPIC_DEFAULT_GEMINI_MODEL": "google-gemini-code",
    "ANTHROPIC_DEFAULT_GPT_MODEL": "openai-gpt-codex",
    "ANTHROPIC_DEFAULT_GLM_MODEL": "glm-glm-coding"
  }
}

This gives me gemini, gpt, and glm as extra model aliases. 9router handles routing each model ID to the correct provider.

Update

claude update

The wrapper intercepts the update command, fetches the latest Claude Code from npm, re-patches it, and reports the version.

Uninstall

curl -sL https://raw.githubusercontent.com/East-rayyy/claude-alias-patch/main/scripts/uninstall.sh | bash

Restores your original Claude Code binary and removes the cache.

What it patches

6 locations in Claude Code's cli.js, all using content-based pattern matching:

#	What	Why
1	Zod enum → string	Accept any model alias, not just sonnet/opus/haiku
2	Env var whitelist	Allow custom env vars through the settings.json loader
3	Model picker fallback	Include custom aliases in the fallback list
4	Tool description	List available aliases in the Task tool's model parameter description
5	Model picker UI	Show custom models in the model selection interface
6	Alias resolver fallback	Resolve custom aliases to their model ID at runtime

All patches are idempotent — running the patcher twice produces the same result.

---

Enjoy! If you run into issues, open an issue or start a discussion.