Alternative LLM Providers

Installation

Part of the claude-code-tools package. See Quick Start for installation.

Overview

You can use Claude Code with alternative LLMs served via Anthropic-compatible APIs. Each provider exposes an endpoint that speaks the Anthropic /v1/messages format, so Claude Code works without any code changes — you just set a few environment variables.

The shell functions below use subshells ( ... ) so the environment variables do not leak into your main shell session. This means you can run multiple instances of Claude Code simultaneously, each using a different LLM.

API Key Setup

Before using any provider, export the corresponding API key in your shell (or add it to your .zshrc / .bashrc):

export KIMI_API_KEY="your-kimi-key"
export Z_API_KEY="your-z-key"
export DEEPSEEK_API_KEY="your-deepseek-key"
export MINIMAX_API_KEY="your-minimax-key"

Shell Functions

Add the function for the provider(s) you want to use to your ~/.zshrc or ~/.bashrc, then use it exactly like the claude command.

Kimi K2

Kimi K2 from Moonshot AI.

kimi() {
    (
        export ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
        export ANTHROPIC_AUTH_TOKEN=$KIMI_API_KEY
        claude "$@"
    )
}

Usage:

kimi                   # start interactive session
kimi -p "explain X"    # one-shot prompt

GLM-4.5 (zai)

GLM-4.5 from Zhipu AI (Z.AI).

zai() {
    (
        export ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic
        export ANTHROPIC_AUTH_TOKEN=$Z_API_KEY
        claude "$@"
    )
}

Usage:

zai                    # start interactive session
zai -p "explain X"     # one-shot prompt

DeepSeek

DeepSeek v3.1 chat model.

dseek() {
    (
        export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
        export ANTHROPIC_AUTH_TOKEN=${DEEPSEEK_API_KEY}
        export ANTHROPIC_MODEL=deepseek-chat
        export ANTHROPIC_SMALL_FAST_MODEL=deepseek-chat
        claude "$@"
    )
}

DeepSeek requires explicitly setting ANTHROPIC_MODEL and ANTHROPIC_SMALL_FAST_MODEL because the default Claude model names are not recognized by the DeepSeek API.

Usage:

dseek                  # start interactive session
dseek -p "explain X"   # one-shot prompt

MiniMax M3.1 Flash / M3

MiniMax M3.1 Flash and M3 from MiniMax.

ccmm() {
    (
        export ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic
        export ANTHROPIC_AUTH_TOKEN=$MINIMAX_API_KEY
        export CLAUDE_CODE_AUTO_COMPACT_WINDOW=524288
        export ANTHROPIC_MODEL='MiniMax-M3.1-Flash[1m]'
        export ANTHROPIC_DEFAULT_SONNET_MODEL='MiniMax-M3.1-Flash[1m]'
        export ANTHROPIC_DEFAULT_OPUS_MODEL='MiniMax-M3.1-Flash[1m]'
        export ANTHROPIC_DEFAULT_HAIKU_MODEL='MiniMax-M3.1-Flash[1m]'
        claude "$@"
    )
}

MiniMax maps the main, Sonnet, Opus, and Haiku model slots to MiniMax-M3.1-Flash[1m] by default. The [1m] suffix declares the one-million-token context window; auto-compaction uses the recommended 512K window (524288). Flash always enables thinking.

Usage:

ccmm                   # start interactive session
ccmm -p "explain X"    # one-shot prompt
ccmm --model 'MiniMax-M3[1m]'  # explicitly select M3

Explicit --model arguments are passed through unchanged, including custom or legacy model IDs. The function does not edit your saved Claude Code settings.

How It Works

Each function:

1. Opens a subshell so env vars are scoped
2. Sets ANTHROPIC_BASE_URL to the provider’s Anthropic-compatible endpoint
3. Sets ANTHROPIC_AUTH_TOKEN to your API key
4. Optionally overrides model names (required for DeepSeek and MiniMax)
5. Launches claude with any arguments you pass

Because everything runs in a subshell, your main shell environment is unaffected. You can have one terminal running kimi and another running zai at the same time.

See Also

• Local LLMs — run models locally with llama.cpp