English · 中文 · GitHub

A collection of launchers and configurators that point Claude Code and Claude Desktop at third-party model backends (DeepSeek, Xiaomi MiMo, etc.).

Author: Agents365-ai · Bilibili

Tools

Tool What it does Platform Backend
dsclaude Claude Code CLI launcher macOS / Linux DeepSeek
mmclaude Claude Code CLI launcher macOS / Linux Xiaomi MiMo
dsclaude-desktop Claude Desktop GUI configurator macOS DeepSeek
dsclaude-desktop.ps1 Claude Desktop GUI configurator Windows DeepSeek
skills/deepseek-vision Vision skill (zero deps) macOS / Linux DashScope Qwen
dsvision-mcp Vision MCP server macOS / Linux DashScope Qwen

Quick start on macOS

git clone https://github.com/Agents365-ai/dsclaude.git
cd dsclaude
chmod +x dsclaude
./dsclaude

dsclaude — Claude Code on DeepSeek

Follows the DeepSeek Anthropic API guide.

export DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxxxx   # add to ~/.zshrc

dsclaude                 # default: deepseek-v4-pro (full reasoning)
dsclaude fast            # deepseek-v4-flash (cheaper / faster)
dsclaude long            # request 1M context window
dsclaude long fast       # 1M + flash

Sets the DeepSeek-recommended env vars (ANTHROPIC_BASE_URL, model mappings, CLAUDE_CODE_EFFORT_LEVEL=max), and exposes the alternate model in Claude Code’s /model picker. Override context window via DSCLAUDE_MAX_TOKENS and effort via DSCLAUDE_EFFORT.

Both models natively support 1M-token context. The [1m] suffix (e.g. deepseek-v4-pro[1m]) is required in Claude Code — dsclaude sets it automatically.

mmclaude — Claude Code on Xiaomi MiMo

export MIMO_API_KEY=sk-xxxxxxxxxxxxxxxxxx       # pay-as-you-go
# or
export MIMO_API_KEY=tp-xxxxxxxxxxxxxxxxxx       # Token Plan

mmclaude                  # start on mimo-v2.5-pro
mmclaude update           # git pull

Auto-detects base URL from the key prefix (sk-* → public, tp-* → Token Plan); override with MIMO_BASE_URL. Sets all four model slots to mimo-v2.5-pro and unsets ANTHROPIC_API_KEY (per MiMo docs).

dsclaude-desktop — Claude Desktop GUI configurator

One-command configurator for Claude Desktop’s built-in Third-Party Inference feature (Developer menu), pre-filled for DeepSeek.

Prerequisites

  1. Claude Desktop installed (claude.ai/download)
  2. Developer Mode enabled (Help → Troubleshooting → Enable Developer Mode, once)
  3. DEEPSEEK_API_KEY environment variable set

Usage (macOS)

export DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxxxx
./dsclaude-desktop        # configure and restart
./dsclaude-desktop -h     # help

Generates an entry under ~/Library/Application Support/Claude-3p/configLibrary/, sets it as appliedId in _meta.json, then restarts the app. Existing GUI-added entries are preserved.

Switching modes

Claude Desktop’s launch chooser handles Anthropic ↔ Gateway switching natively — no --revert flag. Click your profile → Disconnect (or sign out), then pick the other option at next launch.

Classic Chat (claude.ai-style) is unavailable in Gateway mode — it depends on Anthropic-hosted features not exposed via the inference API. Switch back to Anthropic mode to use it.

Usage (Windows)

$env:DEEPSEEK_API_KEY = "sk-xxxxxxxxxxxxxxxxxx"
pwsh ./dsclaude-desktop.ps1

Unlike macOS, Developer Mode is auto-enabled by the script — no manual GUI toggle needed.

Config path: %LOCALAPPDATA%\Claude-3p\configLibrary\ (for Store/MSIX installs, the script also writes to the sandboxed package path as a fallback).

Tested on Windows 11 with Claude Desktop 1.7196 (Windows Store, arm64).

deepseek-vision skill — Vision (zero-dependency)

Gives text-only agents (like DeepSeek) the ability to “see” images. When the agent encounters an image, it calls analyze-image, which sends it to Qwen3.6-Flash and returns a text description.

export DASHSCOPE_API_KEY=sk-xxxxxxxxxxxxxxxxxx
./skills/deepseek-vision/analyze-image /path/to/screenshot.png "What error is shown?"
./skills/deepseek-vision/analyze-image https://example.com/diagram.png

Works with any agent that loads SKILL.md (Claude Code, Cowork, etc.). Default model qwen3.6-flash; override via DSVISION_MODEL and DSVISION_BASE_URL.

Limitation: requires a file path or URL — inline images (drag-drop, paste, “+ → Add files or photos”) aren’t supported. Use dsvision-mcp below for that.

dsvision-mcp — Vision (MCP server)

Same functionality as the skill above, but runs as an MCP server — bypassing two Cowork sandbox limitations:

  1. Network egress — the skill’s DashScope API calls are firewalled inside Cowork’s VM; the MCP server runs outside it
  2. Inline images — auto-picks the latest cached image from ~/.claude/image-cache/, so drag-drop/paste/”+” workflows work (macOS only; Windows Cowork doesn’t cache inline images to disk)

Setup

pip3 install fastmcp requests
export DASHSCOPE_API_KEY=sk-xxxxxxxxxxxxxxxxxx   # add to ~/.zshrc
cd /path/to/dsclaude && pwd    # note the absolute path

Add to the MCP config file matching your mode:

Mode Config file
3P/Gateway (DeepSeek via dsclaude-desktop) ~/Library/Application Support/Claude-3p/claude_desktop_config.json
Standard Anthropic ~/Library/Application Support/Claude/claude_desktop_config.json 
{
  "mcpServers": {
    "dsvision": {
      "command": "/absolute/path/dsclaude/dsvision-mcp"
    }
  }
}

Restart Claude Desktop. The analyze_image tool appears automatically.

Usage

analyze_image()                           # auto: latest cached image
analyze_image(image_path="/abs/path/foo.png")
analyze_image(focus="What error is shown?")

Troubleshooting

Symptom Check
Tool doesn’t appear Wrong config file path / invalid JSON (validate with python3 -m json.tool)
Tool errors DASHSCOPE_API_KEY not set
ModuleNotFoundError Use pip3 not pip
Image not found Pass absolute path, or check ~/.claude/image-cache/ exists

Skill vs MCP: which to use

Scenario Use
Claude Code (CLI), explicit paths skills/deepseek-vision (zero deps)
Cowork / Desktop with inline images dsvision-mcp (only option that works)
Cowork with explicit paths, sandbox tweaks OK either

Community

WeChat Community Group

Support

If these scripts save you time, consider supporting the author:

WeChat Pay
WeChat Pay		 Alipay
Alipay		 Buy Me a Coffee
Buy Me a Coffee		 Give a Reward
Give a Reward

License

MIT