DEV Community

KuaaMU
KuaaMU

Posted on

Give your text-only coding agent eyes — 5-minute setup

#deepseek #ai #programming

DeepSeek V4 Flash writes great code. But it can't see the error dialog youjust pasted, the broken UI, or the screenshot of that rendering bug. It replies"I can't see images." Then you type the error out by hand. Again.

mcp-vision-bridge fixes that. It's a tiny MCP server that gives anytext-only coding agent vision by routing images to a multimodal model youprobably already pay for — mimo, Claude, Gemini, GPT-4o, Qwen-VL. One tool:analyze_image.

You (paste a screenshot of an error dialog)
    ↓
Agent (text-only, calls analyze_image)
    ↓
mcp-vision-bridge (forwards the image bytes)
    ↓
Your vision model (mimo / Claude / Gemini / GPT-4o)
    ↓
Agent now reasons over a full text description
Enter fullscreen mode Exit fullscreen mode

No images ever enter your agent's context. The bridge reads the pixels, yourvision model does the seeing, and the text comes back.

What makes it different

There are plenty of "MCP vision" servers. Most expect you to host a serverthat talks to it. This one is local, one command, zero config:

  • Works across every agent — Claude Code, Codex, opencode, Kimi, PI,Cursor, Reasonix. One server, any MCP client.
  • Paste or drag. Drag an image file into the chat → a real path lands inthe prompt → analyze_image reads it. No clipboard, no paste quirks.
  • Auto-discovery. image="recent" / image="session" find images youpasted across Claude Code, Codex, opencode, Cowork, Reasonix, and Grok —without you typing a path.
  • Many images in one call. Paste 3 screenshots, ask one question. The hookreads your session transcript and passes them all to the model in a singlerequest, each with its own token budget.
  • Current-session only. recent/session never leak images from anothersession — privacy and parallel-use safe.
  • Self-updating. The skill + hook auto-sync into ~/.claude/ on everyserver start, so npx -y mcp-vision-bridge always runs the latest.

Install — pick your agent

Claude Code (plugin, one command):

claude plugin marketplace add KuaaMU/agent-plugins
claude plugin install mcp-vision-bridge
Enter fullscreen mode Exit fullscreen mode

Anything else (Codex, opencode, Kimi, PI…):

git clone https://github.com/KuaaMU/mcp-vision-bridge && cd mcp-vision-bridge
./install.sh            # auto-detects your agent
Enter fullscreen mode Exit fullscreen mode

Manual (any MCP client):

{
  "command": "npx",
  "args": ["-y", "mcp-vision-bridge"],
  "env": {
    "VISION_OPENAI_BASE_URL": "https://your-endpoint/v1",
    "VISION_OPENAI_API_KEY": "sk-your-key",
    "VISION_MODEL": "your-vision-model"
  }
}
Enter fullscreen mode Exit fullscreen mode

Requires Node.js ≥ 18.

Use it

  1. Restart your agent after install.
  2. Drag an image file into the chat (or Ctrl+V in Claude Code / Cowork).
  3. Say "看看这个" / "analyze this" / "what's the error?"

Your agent calls analyze_image and gets back a detailed text description itcan debug against — verbatim text, layout, colors, anomalies.

Pasted several images? The hook detects them in your session transcript andguides the agent to call analyze_image(image="session") — all of them, onecall.

One tool, honest about what it is

analyze_image(image, prompt, task, detail, save_to) accepts a path, URL,"clipboard", "recent", "session", or an array. task presets(describe | ocr | ui | layout | qa) are prompts sent to your vision model— there's no bundled OCR engine; the model does the reading. Pass the user'sactual question as prompt and the model answers what you ask.

What it's not

  • Not a vision model — it routes to one you already pay for.
  • Not an OCR engine — it asks a multimodal model to read the image.
  • Not a cloud service — it runs locally; your keys and images stay on yourmachine.

MIT licensed. npm: mcp-vision-bridge. Source:github.com/KuaaMU/mcp-vision-bridge⭐ If it saves you from transcribing one more error dialog, a star keeps itmaintained.

Top comments (0)