glm-vision MCP Server

Gives any AI Agent the ability to "see" images. Under the hood, it calls 智谱's completely free vision reasoning models, exposes them as standard tools via the MCP protocol, and any MCP-compatible client can use them directly — zero API cost.

💡 Free Guarantee: All free models from the 智谱 benefits zone are used (GLM-4.6V-Flash / GLM-4.1V-Thinking-Flash / GLM-4V-Flash), no paid models are included, and long-term use will not incur any charges.

Exposed Tools

analyze_image(image_input, question, thinking)

Returns: The text analysis result output by the model.

Features

• Local images are automatically read and transmitted as raw base64, no image hosting or public URL required

• Single-image 5MB size limit validation (智谱 restriction)

• Automatic backoff retry 3 times on 429 rate limiting (1.5s / 3s / 4.5s)

• If still failing, automatically degrades: glm-4.6v-flash → glm-4.1v-thinking-flash → glm-4v-flash (all free), and when degraded, the model used is indicated at the beginning of the answer

Related MCP server: Kimi Vision MCP Server

Three-Step Deployment

1. Install Dependencies

pip install -r requirements.txt

⚠️ Key pitfall: mcp must be 1.x. The latest mcp 2.0.0 removed the built-in mcp.server.fastmcp module, which will cause an ImportError if installed. So use pip install "mcp<2".

⚠️ Another pitfall: Do not install zai or zhipuai — on PyPI both names are unrelated placeholder packages (zai belongs to Zalando, and zhipuai's import name has changed). This server uses an OpenAI-compatible client to connect directly to the 智谱 API (base_url=https://open.bigmodel.cn/api/paas/v4); you only need the openai package.

2. Configure API Key

Register at bigmodel.cn with your phone number → create a new one under "API Keys" in the console. Free.

There are two ways to provide the Key, environment variables take higher priority:

• Method A (recommended): Set the environment variable ZHIPU_API_KEY=你的key

• Method B: Rename zhipu_key.txt.example to zhipu_key.txt and fill in the Key (the script will automatically read this file in the same directory).

3. Configure in the Client

Add a block to your client's MCP configuration file (change the path to your own):

{
  "mcpServers": {
    "glm-vision": {
      "command": "python",
      "args": ["/绝对路径/到/glm_vision_mcp.py"],
      "env": {
        "ZHIPU_API_KEY": "你的key"
      }
    }
  }
}

Configuration file locations for each client:

For command, it is recommended to use the absolute path of the Python interpreter (e.g. D:/python/python.exe), to avoid the client not finding python in PATH at startup.

You must restart the client after configuring — the MCP process is spawned only once when the client starts; changing configuration or code will not hot-reload.

Verify It Works

After restarting the client, just have the Agent analyze an image:

Take a look at D:\test.png and tell me what is in this image.

If the Agent says it cannot find the tool, check: whether the configuration file's JSON syntax is valid, whether the script's absolute path is correct, whether the client was restarted, and whether the client requires manually "trusting/enabling" this server (WorkBuddy and Cursor both have this step).

Common Errors Reference

File List