deepseek-claude-proxy — Claude Code at DeepSeek prices

Run Claude Code with DeepSeek's API — same experience, ~50x lower cost

One-line pitch

Replace Anthropic's $15/million tokens with DeepSeek's $0.30/million tokens while keeping the exact same Claude Code workflow.

Cost Comparison

Provider	Price per million tokens	Relative Cost
Anthropic (Claude 3.5 Sonnet)	~$15.00	50x more expensive
DeepSeek (V3.2)	~$0.30	Baseline

That's 50x cheaper. A typical coding session that would cost $1.50 with Anthropic costs just $0.03 with DeepSeek.

Quick Start (3 steps)

1. Install globally

npm install -g deepseek-claude-proxy

2. Run setup wizard

deepseek-claude-proxy init

Follow the prompts to:

• Enter your DeepSeek API key (get one at platform.deepseek.com)
• Configure VSCode settings automatically
• Set up global CLAUDE.md with delegate-first instructions

3. Start coding

deepseek-claude-proxy start

Now open VSCode and use Claude Code as usual — all requests route through DeepSeek automatically.

How It Works

The proxy sits between Claude Code and the AI API:

Claude Code (VSCode) → deepseek-claude-proxy (localhost:1849) → DeepSeek API

Smart Model Routing

• Initial planning turns → deepseek-reasoner (CoT thinking for complex problem decomposition)
• All other turns → deepseek-chat (fast, deterministic coding with "thinking-in-tool-use")

Token Optimization

• Tool results truncated at 12k characters to prevent context explosion
• Output capped at 16k tokens (code responses rarely need more)
• System prompt caching with cache_control: ephemeral to reduce repeated token costs

Claude Compatibility

• Preserves all Claude Code features (tools, streaming, images via text descriptions)
• Handles Anthropic API format exactly
• Automatic fallback to real Claude API if you explicitly request Claude models

Configuration

Environment Variables

DEEPSEEK_API_KEY=sk-...          # Your DeepSeek API key
PROXY_PORT=1849                  # Port to listen on (default: 1849)
PROXY_VERBOSE=true               # Enable verbose logging
ENABLE_VISION=false              # Enable image analysis (experimental)

Config File

Create .deepseek-proxy.json in your project directory:

{
  "apiKey": "sk-...",
  "port": 1849,
  "verbose": false,
  "enableVision": false
}

CLI Commands

# Start proxy server
deepseek-claude-proxy start --port 1849 --key sk-...

# Check status
deepseek-claude-proxy status

# Set API key
deepseek-claude-proxy config set-key sk-...

# Show current config
deepseek-claude-proxy config show

VSCode Integration

Add to your VSCode settings.json:

{
  "claudeCode.environmentVariables": [
    {
      "name": "ANTHROPIC_BASE_URL",
      "value": "http://localhost:1849"
    },
    {
      "name": "ANTHROPIC_DEFAULT_HAIKU_MODEL",
      "value": "deepseek-chat"
    },
    {
      "name": "CLAUDE_CODE_SUBAGENT_MODEL",
      "value": "deepseek-chat"
    }
  ]
}

Or let the setup wizard do it for you:

deepseek-claude-proxy init

Stretching the $20 Claude Pro Subscription

Claude Pro ($20/month) has a rate limit window that resets periodically. In a normal session, every tool call, every file read, every code generation turn burns Anthropic tokens — and long sessions hit the rate limit fast.

This proxy changes that by shifting work to subagents.

When Claude Code uses the Agent tool, it spawns background subprocesses to handle tasks in parallel. With this proxy configured, those subagents inherit the same ANTHROPIC_BASE_URL environment variable and route through DeepSeek — not Anthropic. The main Claude session (your orchestrator) handles high-level reasoning and delegation. The subagents handle the implementation work at DeepSeek prices.

In practice:

• A 2-hour session that would normally exhaust your Claude Pro rate limit might only consume a fraction of it, because most of the token-heavy work (reading files, writing code, running searches) is done by DeepSeek subagents
• The rate limit window stretches from an hour or less to lasting most of the day
• The cost per task drops to near-zero on the DeepSeek side

The included CLAUDE.md template reinforces this by instructing Claude to delegate aggressively — use the Agent tool for anything touching more than one file, run multiple subagents in parallel, and reserve direct execution only for single atomic actions.

The result is that your $20 subscription covers strategic orchestration while DeepSeek handles the labour.

CLAUDE.md Optimization

The proxy includes a pre-configured CLAUDE.md template with:

Delegate-First Architecture

• Use Agent tool for all non-trivial work
• Parallel subagent dispatch (Explore, Executor, Custodian)
• Integration completeness checking

DeepSeek-Specific Guidance

• Model routing strategy
• Token budget awareness
• Tool usage patterns optimized for DeepSeek

Cost Consciousness

• Aggressive parallelization (subagents are cheap!)
• File reading with offset/limit
• Targeted searches over full-file reads

Troubleshooting

"No API key configured"

# Set API key via environment variable
export DEEPSEEK_API_KEY=sk-...

# Or via config file
deepseek-claude-proxy config set-key sk-...

Proxy won't start (port in use)

# Use a different port
deepseek-claude-proxy start --port 1850

# Or find what's using port 1849
lsof -i :1849

Claude Code not connecting

1. Verify proxy is running: deepseek-claude-proxy status
2. Check VSCode settings match proxy port
3. Try health endpoint: curl http://localhost:1849/health

Programmatic Usage

import { startProxy } from 'deepseek-claude-proxy';

const proxy = await startProxy({
  port: 1849,
  apiKey: 'sk-...',
  verbose: true
});

// Proxy runs until stopped
// await proxy.stop();

Development

# Clone and install
git clone https://github.com/yourusername/deepseek-claude-proxy.git
cd deepseek-claude-proxy
npm install

# Development mode
npm run dev

# Build
npm run build

# Test
npm test

License

MIT — see LICENSE file.

Acknowledgements

• DeepSeek for their excellent and affordable API
• Anthropic for the Claude Code ecosystem

---

Save 98% on your Claude Code usage while keeping the exact same workflow.