🤖 Free Claude Code + AUTO_ROUTE

Fork of Alishahryar1/free-claude-code with automatic model routing by task complexity.

🔗 Original project: https://github.com/Alishahryar1/free-claude-code

---

🇬🇧 English | 🇷🇺 Русский

🔥 What it is and why it exists

This is a proxy server for Claude Code CLI. It sits between Claude Code and AI models, and automatically chooses the right model for the task:

If the task is	It goes to	Cost
❓ Simple question / chat	OpenRouter (free models)	💸 $0
⚡ Refactoring / code generation	DeepSeek V4 Flash	⚡ cheap
💪 Complex architecture / algorithms	DeepSeek V4 Pro	💪 powerful

You no longer need to think about which model to use. Just work — the proxy decides for you.

---

🚀 Quick setup (2 minutes)

1. Install Claude Code

npm install -g @anthropic-ai/claude-code

2. Install uv (package manager)

curl -LsSf https://astral.sh/uv/install.sh | sh

3. Clone and install the proxy

git clone https://github.com/smashlight/free-claude-code-AIRouting.git
cd free-claude-code-AIRouting
uv tool install --force .

4. Configure API keys

Create ~/.config/free-claude-code/.env:

mkdir -p ~/.config/free-claude-code

Copy this into it:

# ==== API keys ====
# DeepSeek (required for task classification + flash/pro models)
DEEPSEEK_API_KEY="sk-your-deepseek-key"

# OpenRouter (required for free models on simple tasks)
OPENROUTER_API_KEY="sk-or-v1-your-openrouter-key"

# ==== Models (you can change these) ====
MODEL_OPUS="deepseek/deepseek-v4-pro"           # Most powerful
MODEL_SONNET="deepseek/deepseek-v4-flash"        # Medium tier
MODEL_HAIKU="open_router/openrouter/free"        # Free tier
MODEL="deepseek/deepseek-v4-pro"                 # Fallback

# ==== AUTO_ROUTE ====
AUTO_ROUTE_ENABLED=true
AUTO_ROUTE_CLASSIFIER_MODEL=deepseek/deepseek-v4-flash

# ==== Security ====
ANTHROPIC_AUTH_TOKEN="freecc"

# ==== Thinking ====
# DeepSeek V4 Flash/Pro support thinking; OpenRouter free tier stays off.
ENABLE_MODEL_THINKING=false
ENABLE_HAIKU_THINKING=false
ENABLE_SONNET_THINKING=true
ENABLE_OPUS_THINKING=true

# ==== Optimizations ====
ENABLE_NETWORK_PROBE_MOCK=true
ENABLE_TITLE_GENERATION_SKIP=true
ENABLE_SUGGESTION_MODE_SKIP=true
ENABLE_FILEPATH_EXTRACTION_MOCK=true

5. Start the proxy

In one terminal:

cd free-claude-code-AIRouting
uv run python server.py

6. Create launch aliases

The proxy exposes fcc-claude. Add this to ~/.zshrc (or ~/.bashrc):

alias claudeds='fcc-claude'
alias claudeds-yolo='fcc-claude -- --dangerously-skip-permissions'

Apply it:

source ~/.zshrc

7. Run Claude Code through the proxy

In one terminal, start the proxy:

cd free-claude-code-AIRouting
uv run python server.py

In another terminal, start Claude Code:

claudeds

Done! 🎉 Every request is now automatically routed to the best-fit model.

Important: the regular claude command runs Claude Code directly against Anthropic (official models, paid). Use claudeds to route requests through this proxy with AUTO_ROUTE.

---

🤖 AI-assisted setup (for Windows users)

If you are on Windows or want to automate the setup, give this prompt to an AI assistant (Claude Code, ChatGPT, etc.):

Clone https://github.com/smashlight/free-claude-code-AIRouting.git into a project folder.
Read the README and complete all installation and proxy configuration steps.
Connect it to DeepSeek API using my key: "sk-your-deepseek-key"

IMPORTANT — model names: .env.example contains model names, but DeepSeek may rename models occasionally.
Before starting the proxy, query the DeepSeek API (GET https://api.deepseek.com/models with the API key),
check the current model names (likely deepseek-v4-flash and deepseek-v4-pro), and put them into .env.

Create two aliases:
- claudeds — run Claude Code through the proxy
- claudeds-yolo — same, plus --dangerously-skip-permissions

The assistant can clone the repo, install dependencies, configure .env, check current model names, and create aliases for you.

---

⚙️ How it works

Architecture

                          ┌─ SIMPLE        → OpenRouter (free)
Your request → Proxy ─────┼─ COMPLEX       → DeepSeek Flash
                          └─ VERY_COMPLEX  → DeepSeek Pro
                                ↑
                          Classifier
                         (DeepSeek Flash,
                          ~1.5–2 sec)

Before each request, the proxy makes a fast classification call to DeepSeek Flash (very cheap, about 2 seconds). The classifier looks at the task text and decides its complexity. Then the request is routed to the corresponding model.

Classifier prompt

The classifier receives a prompt like this:

Classify this coding task complexity.
- SIMPLE: Chat, questions, file reads, simple search, one-line fixes
- COMPLEX: Multi-file refactoring, architecture changes, code generation
- VERY_COMPLEX: System design, complex algorithms, performance optimization

Task: {user task text}
Response (one word):

---

🔧 Changing models

The default AUTO_ROUTE setup keeps MODEL_HAIKU on OpenRouter/free without thinking, while MODEL_SONNET (deepseek-v4-flash) and MODEL_OPUS (deepseek-v4-pro) run with thinking enabled. DeepSeek thinking with tool calls depends on replaying prior reasoning_content/thinking blocks in follow-up requests; the DeepSeek adapter preserves that replay and strips unsupported redacted thinking.

The .env file has three tiers:

# Haiku tier — for the simplest tasks
MODEL_HAIKU="open_router/openrouter/free"

# Sonnet tier — for medium tasks
MODEL_SONNET="deepseek/deepseek-v4-flash"

# Opus tier — for complex tasks
MODEL_OPUS="deepseek/deepseek-v4-pro"

# Fallback if the model cannot be resolved
MODEL="deepseek/deepseek-v4-pro"

You can use any model from supported providers:

Provider	Format	Example
DeepSeek	deepseek/model-name	deepseek/deepseek-v4-flash
OpenRouter	open_router/model-name	open_router/anthropic/claude-sonnet-4
NVIDIA NIM	nvidia_nim/model	nvidia_nim/meta/llama-3.3-70b-instruct
LM Studio	lmstudio/model	lmstudio/qwen2.5-7b
Ollama	ollama/model	ollama/llama3.1

---

📋 Requirements

What	Where to get it	Why
DeepSeek API key	https://platform.deepseek.com/api_keys	Task classification + flash/pro models
OpenRouter API key	https://openrouter.ai/keys	Free models for simple tasks
Claude Code CLI	npm install -g @anthropic-ai/claude-code	Main coding interface

---

💡 Examples

Your request	Complexity	Routed to
"What is 2+2?"	SIMPLE	OpenRouter free (free)
"Write a bash script that finds large files"	COMPLEX	DeepSeek Flash
"Design a distributed cache system"	VERY_COMPLEX	DeepSeek Pro
"Refactor auth module to JWT"	COMPLEX → VERY_COMPLEX	DeepSeek Flash → Pro

---

📝 Quick launch commands

# Start the proxy server
cd ~/free-claude-code-AIRouting && uv run python server.py

# Start Claude Code through the proxy (in another terminal)
claudeds

# Fast launch with auto-confirmation
claudeds-yolo

---

🔗 Links

• Original project — free Claude Code proxy
• DeepSeek — DeepSeek API key
• OpenRouter — OpenRouter API key
• DeepSeek-TUI — Auto Mode routing inspiration

---

Made with ❤️ by smashlight