deepseek-mcp

An MCP server that gives Claude Code and Codex a cheap DeepSeek helper for classification, summaries and text cleanup.

  • 56 ★ on the repository
  • Last commit
  • MIT
MCP server

Extended access

Why this level

  • Prompts and any text inside them go to the external DeepSeek API, and the key sits in client settings.
All reasons and checks

arizen-dev/deepseek-mcp

Install

Manual install

DEEPSEEK_API_KEY="sk-..." uvx --from git+https://github.com/arizen-dev/deepseek-mcp.git deepseek-mcp-server

The first README option: run without installing through `uvx`. A persistent install uses `pip install` with the repository address.

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • Prompts and any text inside them go to the external DeepSeek API, and the key sits in client settings.

This is third-party code. Review the repository files before installing.

What it does

`deepseek-mcp` is a small stdio MCP server with two tools. `deepseek` answers quickly and cheaply on a fast non-thinking model. `advise` turns on deep reasoning on the senior model and accepts an `effort` level. The server targets bounded jobs: tag emails, tickets and logs, summarize a document packet, turn messy text into JSON or a table, fill a template, draft mechanical edits. The primary agent reviews the result itself. Architecture decisions, security policy and final client prose stay with a human or the main model.

Who it is for. Developers on Claude Code and Codex who hand mechanical work to a cheap model and review the output themselves.

Good fit when

  • Files, emails or tickets must be sorted into categories with clean JSON output.
  • A pile of notes becomes a summary while the main agent works on something else.
  • A second opinion on a trade-off is needed, so `advise` runs with deep reasoning.
  • Some main-model tokens should be saved on filling templates.

Not a fit when

  • Final client prose goes out without a human read.
  • The hard part of the task is a security or architecture decision.
  • Project content must not be sent to DeepSeek.

Example request

Sort these files into doc, code and config through deepseek and return JSON only.

MCP

Transport
stdio
Authentication
API key

Environment variables

DEEPSEEK_API_KEY required, secret
DeepSeek API key from the platform console.
DEEPSEEK_BASE_URL
Address of another OpenAI-compatible API instead of DeepSeek.

Limitations

The server calls the DeepSeek API, so you need your own key in `DEEPSEEK_API_KEY`, and prompts go to an external service. There is no PyPI package: install from GitHub through `uvx` or `pip`, with Python 3.10 or newer. The worker agent returns text only: it calls no tools, writes no files and cannot see the repository. The author advises treating output as an unchecked draft. Codex does not expand `${DEEPSEEK_API_KEY}` in config, so set the key directly in TOML or through the environment secret mechanism. Restart the client after changing configuration.

How to disable. Remove the `deepseek` entry from the Claude Code MCP configuration or the `mcp_servers.deepseekWorker` section in `~/.codex/config.toml`, then revoke the key in the DeepSeek console.

FAQ

How does `deepseek` differ from `advise`?

`deepseek` answers fast and cheap without reasoning. `advise` turns on deep reasoning on the senior model and suits trade-off analysis.

Can another model be connected?

Yes, any OpenAI-compatible API works. `DEEPSEEK_BASE_URL` sets the address; the README lists examples for Gemini, Ollama, vLLM and Mistral.

How do I see the tools in Claude Code?

After a restart run `/mcp`: the `deepseek` server should appear. The tools are named `mcp__deepseek__deepseek` and `mcp__deepseek__advise`.

Where do I get the API key?

In the DeepSeek console on the keys page. Set it in `DEEPSEEK_API_KEY`; in Claude Code it fits the `env` block of `~/.claude/settings.json`.

Can the server change files?

No. The worker model returns text only and has no tools or repository access.

Discussion

No questions or reviews yet.

An MCP server that runs DeepSeek or your own local model as a sub-agent in Claude Code and Codex with its own tool loop.

MCP serverBroad access59Repository stars

MCP server that connects DeepSeek V4.1 Flash and V4 Pro to Claude Code and Codex with images, FIM and the Responses API

MCP serverExtended access366Repository stars

An MCP server that gives text models sight: it takes a clipboard image, sends it to Qwen-VL and returns a text description.

MCP serverExtended access63Repository stars

A skill that sets DeepSeek up as a native subagent of desktop Codex on macOS and Windows and verifies request routing

SkillExtended access394Repository stars

An MCP server with vector search: the agent describes a task and gets matching skills from the Anthropic and K-Dense collections

MCP serverExtended access411Repository stars
Needs a VPN

An MCP server that hands tasks to the Codex CLI from inside Claude Code: code questions, change review, web search and resumable sessions

MCP serverBroad access649Repository stars

README in short

The server installs from GitHub and connects to the client over stdio. The README example classifies project files. The repository also holds prompt examples, protocol tests and notes on gotchas.

README badge

Report a problem