deepseek MCP server
Commit Oct 15, 2026Exposes DeepSeek models as cheap, supervised workers for Claude Code, Codex, or any MCP-compatible client to handle mechanical tasks.
Also known as arizen-dev/deepseek-mcp
- Stars
- 56
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 15, 2026
- License
- MIT
What is deepseek MCP?
deepseek exposes DeepSeek models as cheap, supervised workers for Claude Code, Codex, or any MCP-compatible client to handle mechanical tasks.
Summary generated from the project's README
About deepseek MCP server
Delegating mechanical, repetitive tasks to a low-cost model is what deepseek-mcp enables for developers using Claude Code, Codex, or other MCP-compatible clients. By acting as a supervised junior analyst, this stdio server provides two specific tools: deepseek for fast, non-thinking flash tasks and advise for deep reasoning pro tasks. It is designed to handle bounded workloads like data extraction, classification, and summarization rather than high-level architectural decisions or autonomous security policy enforcement.
Setup requires a DeepSeek API key provided via environment variables. Users can install the server using uvx, pip, or by cloning the repository directly. The server is compatible with any OpenAI-compatible endpoint, allowing you to point the DEEPSEEK_BASE_URL to local providers like Ollama or other services like Mistral. Every response includes metadata regarding latency, token usage, and estimated cost to help you monitor expenses during mechanical processing tasks.
What you can do with it
- 1Automated data formattingConvert messy or unstructured text into structured JSON or table formats.
- 2Content classificationOrganize large volumes of files, tickets, or logs into specific categories.
- 3Deep reasoning assistanceRequest a second opinion on complex architectural tradeoffs using the reasoning model.
- 4Information summarizationGenerate concise summaries of data packets for human review.
Written from the project's README. Every use case is checked against a quote from it.
Configuration
DEEPSEEK_API_KEY
Automated checks
Ran Nov 13, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablearizen-dev/deepseek-mcp is public and reachablegithub.com/arizen-dev/deepseek-mcp ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-15github.com/arizen-dev/deepseek-mcp ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubdeepseek-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 (the package is installed from GitHub; it is not on PyPI):
DEEPSEEK_API_KEY="sk-..." uvx --from git+https://github.com/arizen-dev/deepseek-mcp.git 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
| Tool | Model | Mode | Best for |
|---|---|---|---|
deepseek |
deepseek-v4-flash | Non-thinking | Classification, extraction, formatting, mechanical edits |
advise |
deepseek-v4-pro | Thinking (effort: medium/high/max) | Architecture, tradeoffs, second opinions, ambiguity |
Cost
Per-call cost depends on token count and model. Pricing per api.deepseek.com (checked 2026-04-30).
| Model | Input (miss) | Input (cache hit) | Output |
|---|---|---|---|
deepseek-v4-flash |
$0.14/1M | $0.0028/1M | $0.28/1M |
deepseek-v4-pro |
$0.435/1M¹ | $0.0036/1M¹ | $0.87/1M¹ |
¹ Pro pricing is 75% off until 2026-05-31. Non-discounted: $1.74/$0.0145/$3.48.
Typical per-call cost (cache miss):
| Task | Flash | Pro |
|---|---|---|
| Small (~1K in + ~0.5K out) | ~$0.0003 | ~$0.0009 |
| Medium (~4K in + ~2K out) | ~$0.001 | ~$0.003 |
Each response footer includes an estimated cost=$... based on token usage.
Environment variables
| Variable | Default | Description |
|---|---|---|
DEEPSEEK_API_KEY |
— | Required. Your DeepSeek API key. |
DEEPSEEK_BASE_URL |
https://api.deepseek.com |
API base URL (change for proxy/compatible providers). |
DEEPSEEK_MCP_LOG |
(unset) | Set to 1 to log call metadata to ~/.deepseek-mcp/calls.jsonl (no prompts logged). |
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:
- reads JSON-RPC messages from stdin;
- exposes two MCP tools:
deepseek(flash, non-thinking) andadvise(pro, thinking); - sends your prompt to DeepSeek's OpenAI-compatible chat completions API;
- streams the response;
- 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:
| Provider | DEEPSEEK_BASE_URL |
Notes |
|---|---|---|
| DeepSeek | https://api.deepseek.com |
Default |
| Google Gemini | https://generativelanguage.googleapis.com/v1beta/openai/ |
Requires Gemini API key; models like gemini-2.5-flash |
| Ollama (local) | http://localhost:11434/v1 |
Run any local model; e.g. llama3, qwen2.5 |
| vLLM (self-hosted) | http://localhost:8000/v1 |
For self-hosted open-weight models |
| Mistral API | https://api.mistral.ai/v1 |
Requires Mistral API key |
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
deepseek MCP questions
Can I use this with local models like Ollama?
Yes, you can use it with any OpenAI-compatible API by setting the DEEPSEEK_BASE_URL environment variable.
Is it safe to use for autonomous decisions?
No, it is designed for bounded tasks and mechanical edits; it is not built for autonomous architecture, security policy, or final judgment.
How do I install it without a permanent installation?
You can use uvx for a zero-install approach directly from GitHub.