dstools — DeepSeek-V4 MCP Toolkit

Give DeepSeek-V4 models eyes and a research desk.

An MCP (Model Context Protocol) server that augments DeepSeek's text models with two capabilities they don't have natively:

1. Image content understanding — DeepSeek-V4 is a text-only model. dstools adds a vision tool that turns any image into rich, structured text the V4 model can reason over (leveraging its 1M-token context and world-class reasoning).

2. Deep Research — a multi-step, citation-backed research pipeline that uses V4 as the planning + synthesis brain over live web search and page extraction.

dstools is a productizable, installable Python package. It speaks MCP over stdio and Streamable HTTP, so any MCP-capable host (Claude Code, Claude Desktop, Cherry Studio, a custom agent, …) can connect a DeepSeek-V4 backend to it and immediately call these tools.

Why this exists

DeepSeek-V4 (deepseek-v4-flash / deepseek-v4-pro, released 2026-04-24) is an outstanding text model with 1M context, strong agentic/tool-calling ability, and an automatic context cache — but the official chat API is text-only (no multimodal vision). dstools closes exactly that gap:

The toolkit is deeply adapted to V4: it defaults to deepseek-v4-pro for synthesis and deepseek-v4-flash for cheap sub-steps, toggles V4's native thinking mode per call, structures prompts for cache hits, and uses V4's JSON-output mode for structured extraction.

Related MCP server: Sibyl

Tools exposed

Granular tools (web_search, fetch_page, analyze_image) let the host agent run its own agentic loop; deep_research is a one-shot orchestrator for when you just want a cited report.

Quick start

# 1. Install (Python 3.10+)
uv sync                # or: pip install -e .

# 2. Configure
cp .env.example .env   # then edit: set DEEPSEEK_API_KEY and a vision provider

# 3. Run the MCP server (stdio — for local hosts like Claude Code/Desktop)
uv run dstools serve

# …or over Streamable HTTP (for remote hosts)
uv run dstools serve --transport http --port 8000

Connect from Claude Code:

claude mcp add --transport stdio dstools -- uv run --directory /path/to/dstools dstools serve

A ready-made examples/claude_desktop_config.json is included for Claude Desktop.

Configuration

All settings are environment variables (.env supported). Sensible defaults mean the keyless parts (search + fetch) work out of the box.

deep_research pipeline (v0.2)

deep_research is a smart, multi-round pipeline (DeepSeek-V4 as the brain):

1. Plan (V4-flash, JSON) → breadth search queries.

2. Round loop (depth rounds): search → fetch → refine — V4-flash reads findings-so-far and generates next-round queries for uncovered facets.

3. Rerank — V4-flash extracts the passages most relevant to the question from each page (always-on; quality over raw stuffing).

4. Synthesize (V4-pro + thinking) → cited markdown report.

Per-step models are tunable; set all RESEARCH_*_MODEL to deepseek-v4-flash for the cheapest runs. dstools doctor prints a per-research cost estimate.

Vision providers (for analyze_image)

Since DeepSeek-V4 can't see images, point VISION_* at any OpenAI-compatible multimodal model:

• OpenAI: VISION_BASE_URL=https://api.openai.com/v1, VISION_MODEL=gpt-4o / gpt-4o-mini

• Alibaba Qwen-VL (DashScope, OpenAI-compat): VISION_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1, VISION_MODEL=qwen-vl-max

• Zhipu GLM-4V: VISION_BASE_URL=https://open.bigmodel.cn/api/paas/v4, VISION_MODEL=glm-4v

• Local (Ollama): VISION_BASE_URL=http://localhost:11434/v1, VISION_MODEL=qwen2.5-vl (no key needed)

Without a vision provider, analyze_image degrades to image metadata + OCR (if pytesseract is installed) and returns a clear note — it never crashes.

Development

uv sync --extra dev
make lint        # ruff
make typecheck   # mypy
make test        # pytest
make serve       # run the server (stdio)

Project layout

src/dstools/
  server.py          # FastMCP server + tool registration
  cli.py             # `dstools` CLI (serve / inspect / doctor)
  config.py          # pydantic-settings config
  llm/               # DeepSeek (OpenAI-compat) + vision clients, V4 thinking-aware
  search/            # pluggable search providers (DuckDuckGo default, Tavily optional)
  web/               # async page fetcher + HTML→Markdown extraction
  tools/             # image / search / fetch / research tools
  utils/             # image I/O & encoding, text chunking
tests/               # pytest suite (network & LLM mocked)
examples/            # claude_desktop_config.json, mcp client demo

License

MIT.