dsclaude — Claude Code & Claude Desktop launchers for alternative backends

中文文档

A small collection of shell scripts that point Claude Code and Claude Desktop at non-Anthropic model backends.

Author: Agents365-ai · Bilibili

Scripts

dsclaude exposes the alternate model in Claude Code's /model picker so you can hot-swap mid-session, sets ANTHROPIC_DEFAULT_HAIKU_MODEL so background/cheap tasks route to the fast model, and honors optional env overrides for context window and output token limits.

dsclaude-desktop is a one-command configurator for Claude Desktop's built-in Configure Third-Party Inference feature (Developer menu). It writes the same config that the dialog would write — pre-filled for DeepSeek — and restarts the app. Claude Desktop's launch chooser handles switching back to Anthropic mode natively, so there's no --revert flag.

Quick start

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

That's it — the bash scripts ship with the executable bit set, so no chmod is needed. Each tool has its own usage section below.

To make dsclaude (Claude Code launcher) globally available:

sudo cp dsclaude /usr/local/bin/

The other tools (dsclaude-desktop, skills/deepseek-vision/analyze-image) reference their own paths or directories, so leave them in the repo.

dsclaude

Follows the official DeepSeek guide: Integrate with Coding Agents / Anthropic API.

export DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxxxx   # add this line to ~/.zshrc or ~/.bashrc

dsclaude                 # start on deepseek-v4-pro (default, full reasoning)
dsclaude fast            # start on deepseek-v4-flash[1m] (cheaper / faster)
dsclaude long            # request a 1M context window (1,048,576 tokens)
dsclaude long fast       # 1M + flash

Sets the DeepSeek-recommended env vars under the hood: ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic, Opus/Sonnet/Haiku model mappings, CLAUDE_CODE_SUBAGENT_MODEL, and CLAUDE_CODE_EFFORT_LEVEL=max (override via DSCLAUDE_EFFORT).

In-session: /model deepseek-v4-flash[1m] ↔ /model deepseek-v4-pro[1m].

Note: Both deepseek-v4-pro and deepseek-v4-flash natively support a 1M-token context window. In Claude Code, the [1m] suffix is required on each model name to enable it (deepseek-v4-pro[1m], deepseek-v4-flash[1m]). dsclaude sets this for you.

mmclaude

Follows the official Xiaomi MiMo Claude Code configuration guide. One model (mimo-v2.5-pro), no fast/long aliases.

export MIMO_API_KEY=sk-xxxxxxxxxxxxxxxxxx       # pay-as-you-go (add to ~/.zshrc)
# or
export MIMO_API_KEY=tp-xxxxxxxxxxxxxxxxxx       # Token Plan

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

The base URL is auto-detected from the key prefix (sk-* → https://api.xiaomimimo.com/anthropic, tp-* → https://token-plan-cn.xiaomimimo.com/anthropic). Token Plan subscribers with a custom subdomain can override with MIMO_BASE_URL=.... Sets all four model slots (main / Opus / Sonnet / Haiku) to mimo-v2.5-pro and unsets ANTHROPIC_API_KEY per the MiMo docs (lingering official credentials shadow the bearer token).

dsclaude-desktop

A one-command configurator for Claude Desktop's built-in third-party inference feature, pre-filled for DeepSeek.

This is not a hack or workaround. Anthropic ships a "Configure Third-Party Inference" dialog inside Claude Desktop (Developer menu) where you can manually point the app at any Anthropic-compatible endpoint. The dialog has six required fields and a model list. dsclaude-desktop writes the same JSON config that the dialog would write, then restarts the app — saving you the menu navigation.

Prerequisites

1. Claude Desktop installed (download from claude.ai/download)
2. Developer Mode enabled in Claude Desktop
   • Help → Troubleshooting → Enable Developer Mode
   • Only needs to be done once. The script verifies this on each run.
3. DeepSeek API Key in $DEEPSEEK_API_KEY, your shell rc, or paste at the prompt

Usage

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

./dsclaude-desktop      # configure and restart Claude Desktop
./dsclaude-desktop -h   # help

What it does:

1. Generates an entry under ~/Library/Application Support/Claude-3p/configLibrary/ with your DeepSeek key, base URL https://api.deepseek.com/anthropic, auth scheme bearer, and deepseek-v4-pro + deepseek-v4-flash (1M context) as the model list
2. Sets appliedId to your entry in _meta.json (existing entries are preserved)
3. Restarts Claude Desktop with killall Claude && open -a Claude

Switching modes

Claude Desktop's launch chooser handles mode switching natively — no --revert flag needed:

Even on the Anthropic sign-in page you can swap back to Gateway:

To switch: click your profile in Claude Desktop → Disconnect (or sign out) → at next launch, pick the other option.

What you get

In Gateway mode the Cowork and Code modes route to DeepSeek. The model picker shows your masked DeepSeek models:

One feature is unavailable: classic Chat (claude.ai-style conversation). Chat depends on Anthropic-hosted features (memory, projects, artifacts, web search) that aren't part of the inference API surface. To use Chat, switch back to Anthropic mode via the launch chooser.

Windows

dsclaude-desktop.ps1 is the PowerShell port. Same JSON schema, same flow:

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

Prerequisites mirror the macOS version: Claude Desktop installed, Developer Mode enabled, DeepSeek API key. The script writes to %APPDATA%\Claude-3p\configLibrary\ instead of ~/Library/Application Support/Claude-3p/configLibrary/.

Untested by the maintainer. The schema and gotchas were discovered on macOS; Anthropic ships the same Electron app on Windows so they should hold, but please open an issue if anything misbehaves.

deepseek-vision skill

A skill that gives any agent (especially text-only ones like DeepSeek) the ability to "see" images. When the agent encounters an image — file path or URL — it calls skills/deepseek-vision/analyze-image, which sends the image to Qwen3.6-Flash (DashScope) and returns a text description the agent can reason over.

export DASHSCOPE_API_KEY=sk-xxxxxxxxxxxxxxxxxx

# Local file:
./skills/deepseek-vision/analyze-image /path/to/screenshot.png "What error is shown?"

# Or an http(s) URL — passed through directly, no download needed:
./skills/deepseek-vision/analyze-image https://example.com/diagram.png

Loaded by any agent that reads SKILL.md files (Claude Code, Cowork, etc.). Default model is qwen3.6-flash; override via DSVISION_MODEL=qwen3.6-plus for higher quality, or DSVISION_BASE_URL=... for a different provider (Xiaomi MiMo-VL via SiliconFlow is one swap away).

Hardening: 10MB image cap with clear error, 60s curl timeout, empty-response detection, exits non-zero with a stderr message on any failure.

Inline-image caveat: this skill needs a file path or URL — it cannot read images that the user drag-drops, pastes, or attaches via Claude Desktop's "+ → Add files or photos" menu. For those use dsvision-mcp below, which runs outside Cowork's sandbox and auto-finds Claude Code's image cache.

In action — Claude Code (CLI) running on DeepSeek via dsclaude:

User pasted a screenshot and said "explain the image". Claude Code recognized the skill (Skill(deepseek-vision) Successfully loaded skill), called analyze-image with the cached path under ~/.claude/image-cache/, and returned an accurate description of the Claude Code startup screen.

dsvision-mcp

A small MCP server that does the same job as the deepseek-vision skill, but bypasses two limitations the skill hits inside Cowork:

1. Sandbox network egress. Cowork's VM only allows outbound traffic to *.anthropic.com / *.claude.com. A bash skill calling dashscope.aliyuncs.com is firewalled. The MCP server runs as a Claude Desktop child process (outside the VM) and bypasses the egress filter.
2. Inline images. Claude Code caches every attached/pasted image to ~/.claude/image-cache/<session-uuid>/N.png on the host filesystem. The MCP server reads from there directly when the agent calls analyze_image() with no path — it auto-picks the most recent cached image. So drag-drop / "+ → Add files or photos" / paste workflows now Just Work. (macOS only. On Windows Cowork, inline images are not cached to disk — pass the file path explicitly instead.)

Prerequisites

• Python 3.8+ (macOS ships Python 3; check with python3 --version)
• DASHSCOPE_API_KEY — get one from DashScope (Alibaba Cloud account required)

Install

# 1. Install Python dependencies
pip3 install fastmcp requests

# 2. Set your API key (add this line to ~/.zshrc so it survives restarts)
export DASHSCOPE_API_KEY=sk-xxxxxxxxxxxxxxxxxx

# 3. Get the absolute path to dsvision-mcp
cd /path/to/dsclaude && pwd    # e.g. /Users/niehu/github/dsclaude

Then add to your Claude Desktop MCP config. Pick the right file for your mode:

If you use both modes, add the entry to both files. If the file doesn't exist yet, create it:

{
  "mcpServers": {
    "dsvision": {
      "command": "/Users/niehu/github/dsclaude/dsvision-mcp"
    }
  }
}

Replace /Users/niehu/github/dsclaude/dsvision-mcp with the path you got from step 3 above.

Restart Claude Desktop. The analyze_image tool will appear to the agent automatically.

Verify it's working

Open Claude Desktop → ask the agent to "analyze the most recent image in cache." If dsvision is connected you'll see it in the Connectors panel (Cowork) or as a tool call in the message stream (Code mode). If the tool doesn't appear, check Claude → Settings → Developer → MCP Logs for startup errors.

Troubleshooting

Usage from the agent's perspective

analyze_image()                     # auto: latest image in ~/.claude/image-cache/
analyze_image(image_path="/abs/path/to/foo.png")
analyze_image(focus="What error is shown?")    # custom prompt

In action — both surfaces of Claude Desktop 3P, running on DeepSeek:

Cowork mode (task agent — dsvision visible in the Connectors panel):

Claude Code in Desktop (Code mode — Used analyze image tool call inline):

In both cases the user attached an image and said "explain the image". DeepSeek agent invoked analyze_image, MCP fetched the cached image from ~/.claude/image-cache/, sent it to Qwen3.6-Flash, and returned an accurate description back into the conversation — including details Qwen read off the image like the model name deepseek-v4-pro[1m] and the working directory.

When to pick which (tldr):

Why a skill instead of an MCP server: zero new dependencies (just bash + curl + jq), no daemon process, single markdown + bash file you can read in 2 minutes.

License

MIT

Support

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

Author

Agents365-ai

• Bilibili: https://space.bilibili.com/441831884
• GitHub: https://github.com/Agents365-ai