deepseek-as-subagent

deepseek-as-subagent

English · 简体中文

      

Run DeepSeek as a real sub-agent inside Claude Code / Codex CLI — not just an LLM endpoint. Claude stays on the main loop (your Max OAuth, your context, your judgment). DeepSeek gets its own Read / Write / Edit / Bash / Glob / Grep / NotebookEdit agent loop for batch / mechanical work.

       Claude (your main agent, Max OAuth, expensive but smart)
         │
         │  decides this is a batch task
         │  →  delegate_to_deepseek(task, context)
         ▼
       DeepSeek sub-agent (cheap, runs its own loop inside the workspace)
         │  Read / Write / Edit / Bash / Glob / Grep / NotebookEdit — all local
         │  iterates until done
         ▼
       Final message bubbles back to Claude
       Claude verifies a sample of the output, reports to you

Quick start

curl -sSL https://raw.githubusercontent.com/PsChina/deepseek-as-subagent/main/curl-install.sh | bash

One line. Clones the repo to ~/.local/share/deepseek-as-subagent, installs the Python package in an isolated venv, registers the MCP server with Claude Code, deploys the skill + /ds slash command, and adds a pure shell alias.

After install, edit ~/.deepseek-mcp/config.json to paste your DeepSeek API key (get one at platform.deepseek.com). Then run claude and try /ds write a python hello world to /tmp/hi.py.

Re-run the same curl | bash later to upgrade. For other clients (Codex, Cursor, Cline) or manual install, see Install below.

How is this different from existing DeepSeek MCP servers?

Most deepseek-mcp-server projects expose DeepSeek as a single LLM call (create_chat_completion, create_anthropic_message). Claude has to read every file itself and feed content into the prompt — DeepSeek only saves you the "thinking" cost, not the "reading/writing" cost. Same for Composio's DeepSeek toolkit.

This project gives DeepSeek its own full agent loop: tool dispatch, file I/O, command execution, multi-turn reasoning — all inside a sandboxed workspace. Claude hands off the entire task and gets a single summary back. The token savings are end-to-end.

What's in the box

• MCP server (Python, stdio transport) exposing one real tool: delegate_to_deepseek(task, context)
• Local agent loop for DeepSeek (agent_loop.py) with OpenAI-compatible function calling
• 7 sandboxed tools for DeepSeek to use: Read / Write / Edit / Bash / Glob / Grep / NotebookEdit (Jupyter cell-level editing)
• Path sandbox + command blacklist (safety.py) — DeepSeek can't escape your workspace or run rm -rf /
• Skill + slash command so Claude knows when to delegate (and when not to)
• pure shell alias for one-shot "no DeepSeek today" runs

Install

Claude Code (default)

git clone https://github.com/PsChina/deepseek-as-subagent
cd deepseek-as-subagent
./install.sh

Then edit ~/.deepseek-mcp/config.json and paste your DeepSeek API key.

Get a key at https://platform.deepseek.com (CNY ¥20 lasts a long time).

Codex CLI

See adapters/codex/ — registers the MCP server with codex and provides instructions you can paste into AGENTS.md.

Cursor / Cline / Claude Desktop / other MCP clients

The MCP server itself is client-agnostic. After pip install -e ., point your client's MCP config at <repo>/.venv/bin/deepseek-mcp. For client-specific "when to delegate" hints, see adapters/ — PRs welcome for new clients.

Usage

After install, just use claude normally. The plugin adds:

• delegate_to_deepseek — Claude auto-invokes it when the task fits (see skills/delegate-to-deepseek/SKILL.md)
• /ds <task> — force delegation, skip Claude's own judgment
• pure shell alias — start Claude with DeepSeek disabled for this session

When delegation actually saves money

Critical rule (encoded in the skill): the delegation decision must happen before Claude reads any source files. If Claude reads first then delegates, both Claude and DeepSeek read the same files — net cost goes up, not down.

The skill enforces: only Glob / LS / read-only Bash are allowed before the delegation decision. If you can't decide without reading, you shouldn't delegate.

Sweet spot for delegation:

• ✅ 10–50 files, mechanical pattern (i18n extract, batch refactor, ETL)
• ✅ Large data + simple processing (log scan, file conversion)
• ❌ Single file < 500 lines (DeepSeek thinking-token overhead > savings)
• ❌ Cross-file design / architectural judgment
• ❌ Tasks needing project-specific idioms from CLAUDE.md

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  Claude Code (CLI or VSCode extension, your Max OAuth)          │
│    ↓ stdio (MCP protocol, no network)                           │
│  deepseek-as-subagent (this project, Python subprocess)         │
│    ↓ HTTPS                                                      │
│  api.deepseek.com (your API key, paid per token)                │
└─────────────────────────────────────────────────────────────────┘

Everything except the actual DeepSeek API call stays on your machine.
No third-party proxy. No cloud relay. Your code never leaves your laptop.

Configuration

~/.deepseek-mcp/config.json:

{
  "api_key": "sk-...",
  "model": "deepseek-v4-pro",
  "max_turns": 50,
  "allowed_tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "NotebookEdit"]
}

Workspace (sandbox root) auto-follows the directory where you launch claude — DeepSeek shares the same scope as Claude itself, no manual config needed. To lock the sandbox to a fixed path regardless of cwd, add "workspace": "/abs/path" to the config.

Override at runtime with env vars: DEEPSEEK_API_KEY, DEEPSEEK_WORKSPACE, DEEPSEEK_MODE=off.

Uninstall

./uninstall.sh

Removes the MCP registration, skill, and slash command. Doesn't touch your Claude Code, your Max OAuth, or your projects.

License

MIT