dsh-free-vision · Free Vision for DSH

English | 中文

A free vision plugin for DeepSeek Harness (dsh): gives text-only models the ability to read images — screenshots, code errors, UI layouts, documents, OCR — powered by free-tier vision models, with a built-in settings UI and zero MCP configuration.

Features

Free-first, multi-provider

Free-tier models work out of the box; switch providers in one click:

Provider	Model	Free quota	Key env var	Signup
Qwen (default)	Qwen3-VL-Flash	Bailian free tier (500K tokens on activation)	DASHSCOPE_API_KEY	Aliyun Bailian
Doubao	Doubao vision	Volcengine free tokens (200K+, up to 500K)	VOLCENGINE_API_KEY	Volcano Ark
SiliconFlow	DeepSeek-OCR	Free OCR calls	SILICONFLOW_API_KEY	SiliconFlow
Zhipu	GLM-4.6V	Metered	ZHIPU_API_KEY	Zhipu
Hunyuan	HY-Vision	Metered	HUNYUAN_API_KEY	Tencent TokenHub
Custom	Any OpenAI-compatible	—	CUSTOM_API_KEY + CUSTOM_BASE_URL + CUSTOM_MODEL_NAME	—

A 1MB screenshot ≈ 2,600 tokens — the Qwen free quota covers ~190,000 images.

Settings UI (Settings → Free Vision)

The plugin registers its own settings section; every option is editable visually and saves take effect immediately (next tool call, no restart):

• Provider cards: click to switch; free tiers carry a "FREE 免费" badge;
• Per-provider key slots: configure once, switching providers automatically uses the matching key;
• Signup links: jump straight to each provider's free-key registration page;
• Live status line: shows the active provider and key source (file / env);
• Advanced settings: model name, tool name, max tokens, temperature, multi-crop, timeout, extra env;
• Graceful no-key mode: the plugin loads fine without a key — a yellow banner and a friendly tool message point to Settings, never an opaque plugin error.

Zero MCP configuration

The vision engine (luma-mcp) ships as a version-locked dependency and runs in-process — no cordis.patch.yml edits, no second "who is actually reading images" channel.

Engineering guarantees

• Auto-reconnect: the engine recovers from crashes with exponential backoff;
• Direct connection: proxy env vars are stripped from the engine process (a stray proxy causes 502);
• One generic tool: image_understand covers OCR / UI / debug / describe tasks;
• CI + 13 unit tests, shipped with every release.

Install

DSH plugins install into a profile (dsh web uses the web profile) via dsh plugin.

Option 1: from npm (recommended)

dsh plugin --profile web add dsh-free-vision

Restart dsh web, then open Settings → Free Vision:

1. Pick a provider card (free tiers carry a badge);
2. Paste that provider's API key into the linkage area (or click the signup link);
3. Click Save — the status line turns green and the model can read images via image_understand.

ERR_PNPM_IGNORED_BUILDS on first install: if pnpm refuses dependency build scripts, add luma-mcp (sharp native dep) to allowBuilds in the profile's pnpm-workspace.yaml and reinstall.

Option 2: from the GitHub repo (development)

git clone https://github.com/FuzzySoul/dsh-free-vision.git
cd dsh-free-vision
npm install
# install into the web profile (Windows cross-drive: use the npm tarball instead of file: paths)

Verify & uninstall

After install, restart dsh web — Settings → Free Vision showing the config cards means it works. You can also run dsh --profile web --dump-config to confirm the plugin layer is mounted.

Uninstall: dsh plugin --profile web remove dsh-free-vision, then restart dsh web.

Technical details: docs/ARCHITECTURE.md.

Configuration

Settings UI vs config file

The UI persists to ~/.dsh/free-vision.json (same fields, UI recommended):

{
  "keys": { "qwen": "sk-...", "volcengine": "" },
  "modelProvider": "qwen",
  "modelName": "",
  "toolName": "image_understand",
  "maxTokens": 8192,
  "temperature": 0.7,
  "multiCrop": true,
  "toolCallTimeoutMs": 200000,
  "lumaEnv": {}
}

Key resolution order: keys[active provider] > legacy top-level apiKey (auto-migrated) > provider env var.

FAQ

Settings say "no API key" but I can read images?

Restart dsh web and check the status line. Early versions had a dual channel (manual mcp-luma + plugin) causing stale state; since 0.5.0 the channel is unified and the UI reflects the actual effective provider and key source.

502 errors?

Bailian & co. are mainland-China endpoints and must be reached directly. The plugin already strips proxy env vars from the engine process; do not set HTTP_PROXY for the plugin yourself.

"Unsupported image format"?

luma detects format by file extension; copy extension-less attachments to .png / .jpg first, or use an http(s) URL / data URI.

How do I switch providers?

Click a card in Settings — each provider keeps its own key and switching takes effect immediately.

Development

npm install
npm test              # vitest unit tests (13)
node test-plugin.mjs  # end-to-end smoke test (needs a provider API key env)

Attribution

Part	Source	License
dsh-free-vision plugin	FuzzySoul	MIT
Vision engine luma-mcp	JochenYang/luma-mcp	MIT
MCP SDK	modelcontextprotocol	MIT

Links

• awesome-dsh-plugin — curated DSH plugin list (this plugin's listing entry);
• dsh-market — visual plugin market inside DSH;
• liustack/modlens — another DSH vision plugin;
• DeepSeek Harness — the official repository.

Star History