Claude Code DeepSeek Delegator

MCP server that lets Claude Code delegate heavy-token tasks to DeepSeek. Claude orchestrates; DeepSeek does the heavy lifting.

One deepseek tool. No spawn. No daemon. Zero dependencies.

─── claude-code-deepseek-delegator
◆ delegated to DeepSeek (v4-pro)

[... full response ...]

─── claude-code-deepseek-delegator · cost ───
deepseek v4-pro $0.073  │  claude sonnet 4 $0.720
saved $0.647 (90%)  │  144,000 tokens (120,000p + 24,000c)
────────────────────────────────

Why not just spawn a subagent?

The usual pattern for heavy tasks — spawning a Claude subagent via the Agent tool — starts a brand new conversation context. That means re-paying the full context window, losing your current state, and still billing at Claude rates.

This MCP server stays in your current session. Claude calls deepseek(...) as a regular tool call: no new context, no re-initialization, no spawn overhead. DeepSeek handles the heavy compute at ~$0.44/M tokens instead of ~$3/M.

The server itself also needs no manual launch. Claude Code starts it automatically via npx when the tool is first called. You configure it once and forget it.

What makes files[] different

When Claude reads files and pastes them into a prompt, those bytes load into Claude's context window first — you pay Claude's rate just to pass content through.

With files[], Claude passes only the paths. The MCP server reads the bytes directly from disk and forwards them to DeepSeek. Large codebases go straight to DeepSeek without ever touching Claude's context.

// Claude calls the tool like this — no Read calls needed first:
deepseek({
  prompt: "Audit these files for security vulnerabilities...",
  files: [
    "/path/to/auth.py",
    "/path/to/middleware.py",
    "/path/to/payments.py"
  ]
})

Install

npm install -g claude-code-deepseek-delegator

Or skip the install entirely — Claude Code will fetch and run it on demand:

{
  "mcpServers": {
    "deepseek": {
      "command": "npx",
      "args": ["claude-code-deepseek-delegator"],
      "env": {
        "DEEPSEEK_API_KEY": "sk-your-key-here"
      }
    }
  }
}

Zero dependencies — Node.js 20+ built-ins only.

Get a free DeepSeek API key at platform.deepseek.com/api_keys. They give you credits on signup.

Configure Claude Code

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "deepseek": {
      "command": "npx",
      "args": ["claude-code-deepseek-delegator"],
      "env": {
        "DEEPSEEK_API_KEY": "sk-your-key-here"
      }
    }
  }
}

Paste your API key directly into the env block. Claude Code does not expand shell variables in mcp.json — the value must be the literal key string.

Restart Claude Code. The deepseek tool is now available.

Automatic delegation

For Claude to automatically suggest delegation before heavy tasks, add the rules from DELEGATION.md to your ~/.claude/CLAUDE.md.

Once added, Claude will ask "Delegate to DeepSeek? (y/n)" before any operation that exceeds ~300 lines, 3+ files, or 500+ words of output.

For stronger enforcement, add the PreToolUse hook from DELEGATION.md to ~/.claude/settings.json — it fires before every skill invocation and injects a mandatory scope-check directly into the model context.

> This task analyzes ~800 lines across 4 files.
> Delegate to DeepSeek? (y/n)

Tools

deepseek

Delegate a task to DeepSeek.

| Param | Type | Default | Description | |-------|------|---------|-------------| | prompt | string | required | The task/instructions for DeepSeek | | files | string[] | — | Absolute file paths to read and include. The MCP server reads them — file contents never pass through Claude's context window. | | system | string | — | Optional system prompt | | model | string | deepseek-v4-pro | Model ID | | temperature | number | 0.7 | 0–2, lower = deterministic | | maxTokens | number | model max | Response token cap |

deepseek_models

List available models with capabilities, context windows, and thinking support.

Models

| Model | Context | Thinking | Best for | |-------|---------|----------|----------| | deepseek-v4-pro | 1M | Yes | Complex analysis, architecture, code | | deepseek-v4-flash | 1M | Optional | Fast/cheap tasks | | deepseek-reasoner | 64K | Yes | Math, logic, step-by-step |

Features

• No spawn — direct tool call, stays in current session, no subagent overhead
• Files bypass Claude — files[] routes disk reads through the MCP process, not Claude's context
• Cost comparison — every response shows DeepSeek cost vs Claude Sonnet 4 equivalent, with $ saved and percentage
• Mandatory delegation gate — Claude can't skip the y/n prompt for heavy tasks
• Retry with backoff — auto-retries on 429/5xx with exponential backoff (configurable attempts)
• Configurable timeout — default 120s, set via DEEPSEEK_TIMEOUT
• Token tracking — prompt, completion, and total tokens on every response
• ANSI colors — output styled to match Claude Code CLI aesthetic (Tokyo Night palette)
• Content-Length framing — MCP spec-compliant JSON-RPC 2.0 over stdio

Environment variables

| Variable | Default | Description | |----------|---------|-------------| | DEEPSEEK_API_KEY | — | Required. Your DeepSeek API key | | DEEPSEEK_API_HOST | api.deepseek.com | API hostname | | DEEPSEEK_TIMEOUT | 120000 | Request timeout in ms | | DEEPSEEK_MAX_RETRIES | 2 | Max retry attempts on failure |

Pricing

| Model | Input (1M tokens) | Output (1M tokens) | |-------|-------------------|---------------------| | deepseek-v4-pro | $0.435 | $0.87 | | deepseek-v4-flash | $0.14 | $0.28 | | deepseek-reasoner | $0.55 | $2.19 | | Claude Sonnet 4 (comparison) | $3.00 | $15.00 |

You get free credits on signup at platform.deepseek.com.

License

MIT