convez.ai
MCP ServerAutomaçãoAvançado
chichuenchang

DeepSeek Worker MCP

chichuenchang/deepseek-worker-mcp

Delega edições em arquivo para DeepSeek barato; Claude orquestra e verifica

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que permite Claude orquestrar tarefas mecânicas de edição em arquivo delegando para um worker DeepSeek mais barato. Claude mantém controle e verificação enquanto DeepSeek executa o trabalho pesado com um conjunto restrito de ferramentas de arquivo (leitura, escrita, edição, glob, grep), sem acesso a shell ou execução de código.

Instalação

$npx convez-ai chichuenchang/deepseek-worker-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Refatoração em massa de arquivos com Claude verificando resultado
  • Processamento automatizado de múltiplos arquivos mantendo segurança
  • Redução de custos delegando tarefas mecânicas e repetitivas a modelo mais econômico

Veredito

Implementação robusta de delegação com modelo seguro (sem shell, sem code execution). Excelente para pipelines onde custo importa e tarefas são mecanicamente estruturadas. Requer cautela com secrets — arquivo de denylist mitiga risco, mas conteúdos lidos vão para API DeepSeek.

README

A tradução para o português deste README ainda está na fila.

deepseek-worker-mcp

An MCP server that lets Claude orchestrate and a DeepSeek worker do bounded, mechanical file work. Claude delegates a sub-task through one tool call; DeepSeek runs its own multi-turn loop over a files-only tool set and returns a structured, verifiable manifest. The point: move bulk token volume to a cheaper model while Claude keeps planning + verification.

This is a clean implementation, informed by an audit of PsChina/deepseek-as-subagent — not a fork.

Security model

The worker has no shell and no code execution — only Read, Write, Edit, Glob, Grep. Because nothing executes, the path jail (resolve() + relative_to(workspace)) is a complete filesystem boundary: there is no write-a-script-then-run-it escape, which is the vector that makes blacklist-based sandboxes leak. Symlinks pointing outside the workspace are filtered.

Residual risk you accept: file contents the worker reads are sent to DeepSeek's API (api.deepseek.com). Don't point it at secrets you don't want transmitted. An optional, empty-by- default denylist lets you refuse secret-pattern paths (e.g. .env*, .git, *.pem).

Install

# Windows
./install.ps1
# macOS / Linux
./install.sh

The installer creates a .venv, installs the package, writes a config template at ~/.deepseek-mcp/config.json (prompts for your DeepSeek API key), registers the server with Claude Code (claude mcp add deepseek -s user), and deploys the delegate-to-deepseek skill. Restart Claude Code afterward to load the new MCP server.

Config — ~/.deepseek-mcp/config.json

{
  "api_key": "sk-...",
  "model": "deepseek-v4-pro",
  "max_turns": 50,
  "workspace": "",
  "allowed_tools": ["Read", "Write", "Edit", "Glob", "Grep"],
  "denylist": [],
  "base_url": "https://api.deepseek.com",
  "thinking": true,
  "reasoning_effort": "max"
}
  • Env overrides: DEEPSEEK_API_KEY, DEEPSEEK_WORKSPACE, DEEPSEEK_MODE.
  • Empty workspace follows Claude Code's launch directory (cwd).
  • allowed_tools is intersected with the files-only set — Bash can never be enabled.
  • thinking enables DeepSeek v4 chain-of-thought (reasoning_content); on by default at reasoning_effort: "max". Effort levels: low/medium/high/max. Set thinking: false or pass thinking="off" per call to save tokens on purely mechanical work.

Usage

In a Claude Code session, just describe a bounded mechanical job; the delegate-to-deepseek skill guides Claude on when to call the tool. Or check health: ask Claude to call ping.

Per-call overrides: delegate_to_deepseek(task, context, model="deepseek-v4-flash", workspace="./src", thinking="on", reasoning_effort="max").

Turning delegation on/off

  • Soft (instruction): tell Claude "do this yourself, don't delegate" / "delegate this to deepseek". Instant, mid-conversation.
  • Hard, on the fly (no restart): ask Claude to "turn deepseek off" / "on" — it calls the deepseek_set_mode tool (on / off / auto). Session-scoped: the override lasts until changed again or the server restarts, then reverts to the launch default. ping shows the effective mode.
  • At launch: start Claude with DEEPSEEK_MODE=off to default the session off.

Result manifest

Every delegation returns the worker's summary plus:

  • files changed (written / edited, with counts)
  • assumptions the worker made
  • could not do (skips/failures)
  • metrics (model, turns, tool calls, tokens, duration)

Always verify by reading a sample of changed files / running tests before declaring success.

Tests

.venv\Scripts\python -m pytest tests/ -q

Offline by default. To run the live smoke test against deepseek-v4-flash:

$env:DEEPSEEK_LIVE = "1"; .venv\Scripts\python -m pytest tests/test_live_smoke.py -q
Ver o original no GitHub

Bom para

Relacionados