English | 简体中文
Vision Bridge gives text-only Agents image understanding through a vision-capable OpenAI-compatible Chat Completions API. It supports OCR, screenshot troubleshooting, chart analysis, and general image understanding.
OpenAI-compatible only describes the API format. The selected model must also support
image_urlinput.
vision-bridge-skill/
├── SKILL.md
├── scripts/
│ └── vision.js
├── references/
│ └── configuration.md
├── test/
│ └── vision.test.js
├── .github/workflows/ci.yml
├── .env.example
├── .gitignore
├── LICENSE
├── package.json
├── README.md
└── README_zh.md
git clone https://github.com/asuojun/vision-bridge-skill.git \
~/.claude/skills/vision-bridgePlace the repository in the Agent's Skill search directory and keep SKILL.md at the repository root. Some Codex environments use:
git clone https://github.com/asuojun/vision-bridge-skill.git \
~/.codex/skills/vision-bridgeCopy the example configuration:
cp .env.example .envSet your provider details:
OPENAI_API_KEY=your-api-key
OPENAI_BASE_URL=https://provider.example/v1
VISION_MODEL=your-vision-capable-modelFor a no-auth local service, omit OPENAI_API_KEY:
OPENAI_BASE_URL=http://127.0.0.1:8000/v1
VISION_MODEL=local-vlmThe script only loads .env from the Skill root automatically. To use another file, set it explicitly:
export VISION_ENV_FILE=/absolute/path/to/vision.envSee references/configuration.md for all options.
Analyze a local screenshot:
node scripts/vision.js ./screenshot.png \
"Identify the visible error and suggest troubleshooting steps"Analyze a remote image:
node scripts/vision.js https://example.com/chart.png \
"Summarize the chart's trends and anomalies"Run high-detail OCR:
node scripts/vision.js --detail high ./scan.jpg \
--prompt "Transcribe every visible line and mark uncertain characters as [?]"Return JSON:
node scripts/vision.js --json ./image.png "Describe the image"Show all CLI options:
node scripts/vision.js --helpThe provider must support:
POST /chat/completions- OpenAI-style
modelandmessages - array-form
messages[].content textandimage_urlcontent parts- image URLs or base64 data URLs
- responses at
choices[0].message.content
Some models use max_completion_tokens instead of max_tokens:
export VISION_MAX_TOKENS_FIELD=max_completion_tokensIf the provider rejects the optional detail field:
export VISION_DETAIL=noneImages are sent to the configured vision API. Before sending identity documents, credential screenshots, customer data, internal systems, or other sensitive content, confirm that the user has authorized the provider to process that data.
The script does not scan directories and does not automatically load .env files from the current working directory. Never commit API keys or include them in Issues, logs, or screenshots.
Node.js 18 or newer is required:
npm install
npm run ciTests use a local mock server and never call a real vision API.