DMontgomery40/deepseek-mcp-server

DeepSeek MCP Server – What It Is

DeepSeek MCP Server is a small Node‑JS program that acts as a bridge between the Model Context Protocol (MCP) and DeepSeek’s V4.1 Flash and V4 Pro language‑model APIs. In practice it lets tools that understand MCP (for example Claude, Codex, or custom MCP‑aware clients) talk to DeepSeek without each client having to implement DeepSeek’s own HTTP API.

Core capabilities

Feature	How it works	What you get
Chat completions	Sends a `/v1/chat/completions` request (or the beta endpoint) to DeepSeek, optionally with images.	Text replies, JSON‑structured output, streaming, and optional in‑memory conversation memory (`conversation_id`).
Responses API	Calls DeepSeek’s Responses endpoint, which can return text or images.	Stateless calls that can include custom tool `apply_patch` and structured output.
FIM completion	Uses DeepSeek’s `/beta/completions` endpoint for fill‑in‑the‑middle generation.	Text completions with optional streaming, limited to ~4 K output tokens.
Model discovery & balance	Wraps DeepSeek’s `/models` and `/balance` endpoints.	Lists available models, their limits, and shows your account credit.
Files API	Uploads, lists, retrieves and deletes image files on DeepSeek’s storage.	Re‑usable `file_id`s for vision inputs, with size limits (64 MiB per file).
Conversation memory	Keeps a short‑lived, in‑process store of recent messages per `conversation_id`.	Ability to continue a chat without resending the full history each request.
Rate‑limit & auth	Simple token‑based auth for the optional HTTP transport and a global tool‑rate‑limit.	Prevents accidental over‑use of your DeepSeek quota.

How you run it

STDIO mode (default) – launch with npx deepseek-mcp-server@1 and pipe MCP JSON messages over stdin/stdout. Ideal for local CLI tools.
HTTP mode – set MCP_TRANSPORT=streamable-http and specify host/port. The server then listens at http://127.0.0.1:3001/mcp and can be called from browsers or other services.

Both modes require only one environment variable: DEEPSEEK_API_KEY.

Who would use this?

Developers building MCP‑aware assistants (Claude, Codex, etc.) who want DeepSeek as a backend without writing DeepSeek‑specific request code.
RAG or multimodal pipelines that need to send images to DeepSeek’s vision model; the server handles the file‑upload lifecycle for you.
Testing or prototyping – the npm package ships with a live‑smoke‑test suite that exercises every tool against a real DeepSeek account.

What it does not do

It never generates images, video, or audio – only text and image understanding.
It is not a full‑featured gateway for arbitrary OpenAI‑compatible APIs; it only implements the endpoints DeepSeek currently offers.

Quick start (copy‑paste)

# 1️⃣ Install/run (Node 20+ required)
DEEPSEEK_API_KEY=YOUR_KEY npx -y deepseek-mcp-server@1

# 2️⃣ Example MCP client config (JSON)
{
  "mcpServers": {
    "deepseek": {
      "command": "npx",
      "args": ["-y", "deepseek-mcp-server@1"],
      "env": {"DEEPSEEK_API_KEY": "YOUR_KEY"}
    }
  }
}

Now any MCP‑compatible client can call tools like chat_completion, list_models, or upload_file through the deepseek server.

Where to learn more

Official DeepSeek docs – links in the README for model limits, pricing, and vision guide.
MCP specification (2026‑07‑28) – the protocol this server implements.
GitHub repo – contains the full source, tests, and audit documents (docs/research/...).

License

MIT – you can freely use, modify, and redistribute the server.