deepseek-mcp

Use DeepSeek from Claude Code, Codex, or any MCP-compatible client as a small, cheap supervised worker.

deepseek-mcp is a tiny stdio MCP server with two tools:

deepseek(prompt, system?)     — fast, cheap, non-thinking (flash)
advise(prompt, system?, effort?) — deep reasoning (pro + thinking)

It is built for bounded tasks where another model can reduce mechanical load:

• classify inboxes, tickets, logs, notes, or docs;
• summarize packets;
• turn messy text into JSON or tables;
• populate templates;
• generate first-pass mechanical edits;
• produce reviewable candidate output for a human or primary agent.

It is not built for autonomous architecture, security policy, final client prose, or decisions where the hard part is judgment.

Quickstart

1. Install

Zero-install with uvx:

DEEPSEEK_API_KEY="sk-..." uvx deepseek-mcp-server

Or install persistently:

pip install "git+https://github.com/arizen-dev/deepseek-mcp.git"

Or clone and install locally:

git clone https://github.com/arizen-dev/deepseek-mcp.git
cd deepseek-mcp
pip install -e .

2. Add your DeepSeek key

Create an API key at:

https://platform.deepseek.com/api_keys

Then export it:

export DEEPSEEK_API_KEY="sk-..."

For Claude Code, the lowest-friction setup is to put the key in your global settings:

{
  "env": {
    "DEEPSEEK_API_KEY": "sk-..."
  }
}

Path:

~/.claude/settings.json

MCP servers are started when the client launches, so restart Claude Code or Codex after changing env/config.

3. Configure MCP

If you installed via pip (or uvx), use the installed command directly:

{
  "mcpServers": {
    "deepseek": {
      "command": "deepseek-mcp-server",
      "args": [],
      "env": {
        "DEEPSEEK_API_KEY": "${DEEPSEEK_API_KEY}"
      }
    }
  }
}

If you cloned the repo, point to the script directly:

{
  "mcpServers": {
    "deepseek": {
      "command": "python3",
      "args": ["/absolute/path/to/deepseek-mcp/deepseek_mcp_server.py"],
      "env": {
        "DEEPSEEK_API_KEY": "${DEEPSEEK_API_KEY}"
      }
    }
  }
}

After restart, /mcp should show a deepseek server.

In Claude Code, the tool names are:

mcp__deepseek__deepseek     — flash (fast, mechanical)
mcp__deepseek__advise       — pro  (deep reasoning)

Codex

For Codex, add a global MCP server in ~/.codex/config.toml:

[mcp_servers.deepseekWorker]
command = "deepseek-mcp-server"
args = []

[mcp_servers.deepseekWorker.env]
DEEPSEEK_API_KEY = "sk-..."

Codex TOML does not expand "${DEEPSEEK_API_KEY}" in the same way Claude project MCP configs do. Put the key directly in the TOML env block or use whatever secret mechanism your Codex environment supports.

Demo

Prompt:

Classify these files into doc / code / config. Return JSON only:
- README.md
- pyproject.toml
- src/deepseek_mcp/server.py

Example output:

[
  {"file": "README.md", "type": "doc"},
  {"file": "pyproject.toml", "type": "config"},
  {"file": "src/deepseek_mcp/server.py", "type": "code"}
]

The server appends lightweight metadata:

---
_deepseek · model=deepseek-v4-flash  latency=18.42s  tokens=52+74  cost=$0.0001_

Latency depends heavily on prompt size, model, network, and API load. Treat benchmark numbers as directional, not a guarantee.

CLI

After installing, you can use the CLI for smoke tests and one-shot calls:

# Validate setup
python -m deepseek_mcp check

# One-shot flash call
python -m deepseek_mcp run "Classify: urgent / later — 'Server down in prod'"

# Advisor call with deep reasoning
python -m deepseek_mcp advise "Should we build or buy analytics?" --effort max

Exit codes: 0 = success, 1 = API error, 2 = missing key.

Models

Cost

Per-call cost depends on token count and model. Pricing per api.deepseek.com (checked 2026-04-30).

¹ Pro pricing is 75% off until 2026-05-31. Non-discounted: $1.74/$0.0145/$3.48.

Typical per-call cost (cache miss):

Each response footer includes an estimated cost=$... based on token usage.

Environment variables

When to use it

Good:

• "Classify these 200 filenames. Mark uncertainty."
• "Turn this rough note into a CSV table."
• "Extract all TODOs and group them by owner."
• "Create candidate JSON from this messy list. Use null for missing values."
• "Summarize this packet for review; do not make decisions."

Bad:

• "Design my architecture."
• "Write the final client email."
• "Decide whether this is secure."
• "Resolve this ambiguous business rule."
• "Publish this reply directly."

Use it like a fast junior analyst whose work you will review, not like an owner.

How it works

The server:

1. reads JSON-RPC messages from stdin;
2. exposes two MCP tools: deepseek (flash, non-thinking) and advise (pro, thinking);
3. sends your prompt to DeepSeek's OpenAI-compatible chat completions API;
4. streams the response;
5. returns the text plus model, latency, token, and cost metadata.

There is no database, no background daemon, no local web server, and no file-system access beyond the MCP client starting the process.

Smoke test

After installing:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \
  | DEEPSEEK_API_KEY="sk-..." deepseek-mcp-server

You should see a JSON response with two tools (deepseek + advise).

Then test a real call:

echo '{"jsonrpc":"2.0","id":2,"method":"tools/call",\
  "params":{"name":"deepseek","arguments":{"prompt":"Return exactly: ok"}}}' \
  | DEEPSEEK_API_KEY="sk-..." deepseek-mcp-server

Examples

See examples/ for real prompt templates:

• flash_classify.md — inbox triage
• advise_architecture.md — architecture decision
• advise_tradeoff.md — build vs buy

Benchmark

See docs/benchmark.md for validation observations and usage guidance.

Development

pip install -e ".[dev]"
pip install -r requirements-dev.txt  # alternative
python -m pytest

Compatible endpoints

deepseek-mcp works with any OpenAI-compatible API. Set DEEPSEEK_BASE_URL to point elsewhere:

Security notes

• The worker returns text only. It cannot call tools, write files, or access your repo. Output lands in the primary model's context — you review before anything is used.
• Do not commit API keys.
• Prefer client/global env injection over hardcoding keys in project repos.
• Treat model output as untrusted candidate text.
• Do not give the tool access to secrets you would not paste into DeepSeek directly.
• Review output before it reaches users, customers, production systems, or public channels.

License

MIT