image-analysis-fallback

中文文档

Bridge any non-multimodal model to an external vision API: when the current model cannot see an image, this skill sends it to a multimodal model (default: qwen3.7-plus) and feeds the textual result back, so the agent can continue the task.

Works with any agent that supports SKILL.md — Claude Code, Codex, Cursor, ZCode, and more. No platform-specific dependencies.

Install

Clone or symlink this repo into any skill directory:

git clone https://github.com/anghunk/image-analysis-fallback.git ~/.claude/skills/image-analysis-fallback

Tool	Skill directory
Claude Code	~/.claude/skills/
Codex	~/.codex/skills/
Cursor	.cursor/skills/
ZCode	~/.zcode/skills/

Configuration

Env vars take precedence over config.json:

# Option 1: config file
cp config.example.json config.json   # fill in your api_key

# Option 2: env vars
export VISION_API_KEY=sk-xxxx

Usage

The skill triggers automatically in new sessions when the user provides an image the current model cannot see. You can also call it manually:

python3 analyze_image.py <image-path> "[question]"

Supported API formats

Set api_type in config.json (or VISION_API_TYPE) to choose the API dialect:

api_type	Providers	Example api_url
anthropic (default)	Claude, DashScope Anthropic-compatible	https://api.anthropic.com/v1/messages
openai	OpenAI, Qwen OpenAI-compatible mode, Ollama, vLLM, LM Studio	https://api.openai.com/v1/chat/completions
gemini	Google Gemini native (api_url is the base URL; model name is appended)	https://generativelanguage.googleapis.com/v1beta
dashscope	Alibaba DashScope native multimodal API	https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation

Options

Env var	Description	Default
VISION_API_TYPE	API dialect: anthropic / openai / gemini / dashscope	anthropic
VISION_API_KEY	API key (overrides config.json)	from config.json
VISION_API_URL	API endpoint (for gemini: base URL)	per api_type
VISION_MODEL	Model name	qwen3.7-plus
VISION_MAX_TOKENS	Max output tokens	1024