DeepCode V4

🇪🇸 Leer en Español

DeepSeek V4 Pro + Claude Code + Local Vision. A single command. The ultimate proxy that makes DeepSeek work seamlessly as if it were Claude — with native tool-calling, 1M context tokens, and image support via local LLMs. All for a fraction of the cost.

✨ Why DeepCode?

• 👁️ Vision for DeepSeek — DeepSeek doesn't support images. DeepCode does. It auto-detects Ollama or LM Studio and gives DeepSeek eyes using a local vision model. Zero configuration required.
• 🧠 Native Thinking Mode Support — Renders DeepSeek's reasoning (R1) directly into Claude Code's collapsible ∴ Thinking… block. Supports chained tool-calls without losing context.
• 🎛️ Native Effort Control — Use the /effort low or /effort max commands in the console; the proxy translates Anthropic's tokens to DeepSeek's engine in real-time.
• 🔧 100% compatible with Claude Code — Same flags, same tools, same environment. Use deepcode exactly as you would use claude.
• 💰 95% cheaper — DeepSeek V4 Pro costs ~$0.04 for every $0.90 you would spend on Claude Sonnet.
• 🔄 Continue Claude sessions — Ran out of tokens? Run deepcode --resume and pick up right where you left off.
• 📊 Real-time Statusline — Track tokens consumed and direct costs right in the bottom bar of Claude Code.

📦 Prerequisites

Before installing DeepCode, make sure you have:

1. Claude Code installed and working on your system. You can install it following the official Anthropic guide.
2. DeepSeek V4 API Key — Get your key at platform.deepseek.com. DeepCode uses this API key to route requests to DeepSeek V4 Pro.

Important

Without Claude Code installed, deepcode will not run. Without a valid DeepSeek API key, requests cannot be processed.

🚀 Installation

npm install -g deepcode-v4

On the first launch, if no DEEPSEEK_API_KEY is found, DeepCode prompts for it in the console (masked input), validates the format (sk-...), and saves it globally to ~/.deepcode-v4/.env. You only do this once — works across all projects, no per-project .env required.

deepcode            # First run → prompts for API key, then starts
deepcode --setup    # Reconfigure / replace the saved API key

Get your key at platform.deepseek.com/api_keys.

📋 Usage

deepcode                              # New session
deepcode "create an Express REST API" # Direct prompt
deepcode --setup                      # Configure or replace global DEEPSEEK_API_KEY
deepcode --no-vision                  # DeepCode specific: Disable auto-detected vision

# --- Native Claude Code Flags ---
# DeepCode acts as a transparent wrapper, so all native Claude flags work flawlessly:
deepcode --resume                     # Open interactive picker to resume a previous session
deepcode --resume <session-id>        # Continue a specific previous session directly
deepcode --dangerously-skip-permissions  # Autonomous mode (auto-approve all tool uses)

Any other claude flag works perfectly with deepcode.

Native in-console commands:

• /effort low or /effort max: Control DeepSeek's thinking depth and time on the fly.
• /model sonnet: Instantly switch to the ultra-fast deepseek-v4-flash version.

⚙️ Configuration

The interactive setup at first launch covers the common case. For advanced setups, DEEPSEEK_API_KEY is resolved in this precedence order:

1. Shell environment variable (export DEEPSEEK_API_KEY=sk-...)
2. .env file walking up from the current working directory (per-project override)
3. ~/.deepcode-v4/.env (global, written by deepcode --setup)

Run deepcode --setup anytime to replace the saved global key. Check .env.example for all optional vars (model overrides, retry, timeouts, telemetry).

⚠️ Security: Your API key stays on your local machine. The global file ~/.deepcode-v4/.env is created with 0600 permissions on POSIX. If you put a .env inside a project, remember to add .env to your .gitignore.

👁️ Local Vision (Auto-detected)

DeepSeek cannot see images. DeepCode gives it eyes.

When you run deepcode, the proxy automatically detects if you have a local vision LLM running. If it finds one, it intercepts images, converts them into highly detailed textual descriptions, and passes them to DeepSeek so it can reason about them. If no local LLM is found, it falls back to normal text mode.

Enable vision (3 steps):

1. Install Ollama
2. Run ollama pull qwen2-vl:7b
3. Run deepcode — vision is automatically activated!

Recommended Models:

Model	VRAM	Quality
qwen2-vl:7b	~5 GB	Very Good
qwen3-vl:8b	~6 GB	Excellent
llava:7b	~5 GB	Good
moondream	~2 GB	Lightweight

It also supports LM Studio (localhost:1234).

📜 License

MIT