Claude Code

Claude Code is Anthropic’s agentic coding tool that lives in your terminal. It can understand your codebase, edit files, run commands, and handle complex development tasks. By pointing it at Hugging Face Inference Providers, you can use open models like GLM-5.1, Gemma 4, and more as the backing LLM.

Overview

Claude Code supports custom API endpoints via environment variables. By setting ANTHROPIC_BASE_URL to the Hugging Face router and providing your HF token, all Claude Code requests are routed through Inference Providers, giving you access to a wide range of open models.

Prerequisites

Claude Code CLI installed (installation guide)
A Hugging Face account with API token (needs “Make calls to Inference Providers” permission)

Configuration

Option 1: Using the hf-claude Extension (Recommended)

The hf-claude extension for the hf CLI provides an interactive model and provider picker that launches Claude Code with the right environment variables preconfigured.

Make sure you have HF_TOKEN set:

export HF_TOKEN="hf_..."

Install the extension:

hf extensions install hanouticelina/hf-claude

Launch Claude Code through hf:

hf claude

The extension will present an interactive picker where you can select a model from the available Inference Providers models, and optionally choose a specific provider. It then launches Claude Code with all the necessary environment variables set.

You can also forward any extra arguments to Claude Code:

hf claude --help

Option 2: Manual Environment Variables

Set the following environment variables before launching Claude Code:

export ANTHROPIC_BASE_URL="https://router.huggingface.co"
export ANTHROPIC_AUTH_TOKEN="${HF_TOKEN}"
export ANTHROPIC_API_KEY="${HF_TOKEN}"
export ANTHROPIC_DEFAULT_OPUS_MODEL="zai-org/GLM-5.1"
export ANTHROPIC_DEFAULT_SONNET_MODEL="zai-org/GLM-5.1"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="zai-org/GLM-5.1"
export CLAUDE_CODE_SUBAGENT_MODEL="zai-org/GLM-5.1"

Then start Claude Code:

claude

Replace zai-org/GLM-5.1 with any model available on Inference Providers. You can also append a provider suffix to pin a specific provider (e.g. MiniMaxAI/MiniMax-M2.7:fireworks-ai). Setting an explicit provider gives you deterministic routing, while omitting it (e.g. MiniMaxAI/MiniMax-M2.7) lets the router fall back to alternative providers for better resilience.

You can also append a :cheapest or :fastest suffix to prefer cheaper or faster providers (e.g. MiniMaxAI/MiniMax-M2.7:cheapest).

The ANTHROPIC_DEFAULT_*_MODEL variables map to Claude Code’s internal model slots (Opus, Sonnet, Haiku), from the most powerful to the quickest. You can assign different models to each slot to balance capability and speed e.g. zai-org/GLM-5.1 for Opus, google/gemma-4-31B-it:together for Sonnet, and openai/gpt-oss-120b:cerebras for Haiku. CLAUDE_CODE_SUBAGENT_MODEL controls which model is used for sub-agents.

Resources

Update on GitHub