multimodal-mcp-server

Vision / multimodal MCP server that talks to any OpenAI-compatible gateway (/v1/chat/completions). Built for reasoning models: thinking is off by default and there is a reasoning_content fallback, so models like mimo-v2.5 return real answers instead of Invalid API response: missing content.

Exposes 8 tools: analyze_image, extract_text_from_screenshot, ui_diff_check, understand_technical_diagram, analyze_data_visualization, diagnose_error_screenshot, ui_to_artifact, analyze_video.

Why this fork-style patch exists

Most vision MCP servers hard-code thinking: { type: 'enabled' } and read only message.content. Reasoning models return content: null while thinking, so those servers throw Invalid API response: missing content. This package:

1. Turns thinking off by default — controlled by THINKING (default disabled), so reasoning models answer in content directly.
2. Falls back to reasoning_content — if content is empty, it reads message.reasoning_content (belt and braces).

Quick start (Claude Code)

{
  "mcpServers": {
    "multimodal-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "[email protected]"],
      "env": {
        "BASE_URL": "http://127.0.0.1:3000/v1/",
        "API_KEY": "sk-your-gateway-token",
        "VISION_MODEL": "mimo-v2.5",
        "THINKING": "disabled"
      }
    }
  }
}

Restart Claude Code → /mcp shows multimodal-mcp-server online → ask it to analyze_image a picture.

You can also run it directly: npx -y multimodal-mcp-server (needs the same env vars).

Env vars

| Var | Required | Notes | |---|---|---| | BASE_URL | yes | OpenAI-compatible base URL, must end with / (the server concatenates BASE_URL + 'chat/completions' literally) | | API_KEY | yes | gateway token. MUST be set explicitly — some shells inject ANTHROPIC_AUTH_TOKEN, which would otherwise 401 against your gateway | | VISION_MODEL | yes | vision model id, e.g. mimo-v2.5 | | THINKING | no | enabled/disabled, default disabled. Keep disabled for reasoning models |

Optional: TIMEOUT (default 300000ms), RETRY_COUNT, VISION_MODEL_TEMPERATURE / _TOP_P / _MAX_TOKENS. Log file path: MCP_LOG_PATH.

The patch

The server is based on an Apache-2.0 MCP vision package (v0.1.4), patched in build/core/chat-service.js:

-            thinking: { type: 'enabled' },
+            thinking: { type: process.env.THINKING === 'enabled' ? 'enabled' : 'disabled' },
...
-            const result = response.choices?.[0]?.message?.content;
+            const message = response.choices?.[0]?.message;
+            const result = message?.content || message?.reasoning_content;

License

Apache-2.0. Upstream: @z_ai/mcp-server 0.1.4 (original author attribution retained).