MCP 服务器

Kimi Vision MCP Server

Enables analysis of local images through Kimi (Moonshot AI) vision models via the MCP protocol, supporting features like OCR and long context understanding.

README

Kimi Vision MCP Server

MCP server for Kimi (Moonshot AI) vision models — analyze images through the OpenAI-compatible /chat/completions endpoint at https://api.moonshot.cn/v1.

Features

🖼️ Analyze images (local files: png/jpg/jpeg/gif/webp)
🧠 Auto-pick strongest thinking mode per model:
- kimi-k3 → reasoning_effort: "max"
- kimi-k2.6 / kimi-k2.5 / kimi-k2.7-code → thinking: {type: "enabled"}
📏 Up to 1M token context (kimi-k3)
💰 Same billing as Moonshot V1 — see pricing
⚡ One-line npx deploy, zero non-MCP dependencies

Requirements

Node.js >= 18
A platform.kimi.com API key

Install & run

cd kimi-vision-mcp-server
npm install
npm start

Environment variables

Variable	Required	Default	Description
`KIMI_API_KEY`	✅	—	Your Kimi/Moonshot API key. (`MOONSHOT_API_KEY` also accepted.)
`KIMI_MODEL`		`kimi-k3`	Model name. Vision-capable: `kimi-k3`, `kimi-k2.7-code`, `kimi-k2.6`, `kimi-k2.5`, `moonshot-v1-*-vision-preview`.
`KIMI_BASE_URL`		`https://api.moonshot.cn/v1`	Override endpoint (for proxies).
`KIMI_MAX_TOKENS`		`4096`	Default max output tokens.

Claude Code / CC-Switch config

{
  "mcpServers": {
    "kimi-vision": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "kimi-vision-mcp-server"],
      "env": {
        "KIMI_API_KEY": "your-kimi-api-key",
        "KIMI_MODEL": "kimi-k3"
      }
    }
  }
}

Or run from a local checkout:

{
  "mcpServers": {
    "kimi-vision": {
      "command": "node",
      "args": ["D:\\GitHub\\Claude\\MCP\\Kimi\\kimi-vision-mcp-server\\src\\index.js"],
      "env": {
        "KIMI_API_KEY": "your-key-here"
      }
    }
  }
}

Tool: `kimi_vision_understand`

Parameter	Type	Required	Description
`image`	string	✅	Local image file path (`C:/path/to/x.png`) or `ms://<file-id>` for pre-uploaded. Remote HTTP URLs NOT supported.
`prompt`	string	✅	What to ask about the image.
`detail`	enum		`auto` / `low` / `high`. Default `auto`.
`max_tokens`	number		Max output tokens. Default 4096.
`thinking`	bool		Enable reasoning. Default `false`. Auto-mapped per model.

Important: Remote URLs not supported

Kimi's vision API does not accept remote HTTP image URLs (per the official docs). You must either:

Pass a local file path (the MCP will inline it as base64), or
Upload to Moonshot first and pass ms://<file-id> (advanced).

Remote URLs are rejected with a clear error message.

Why no `temperature` parameter?

Per the Kimi API model params reference, all current Kimi models have fixed temperature:

Model	Temperature
`kimi-k3`	fixed 1.0
`kimi-k2.7-code` (and -highspeed)	fixed 1.0
`kimi-k2.6` thinking	fixed 1.0
`kimi-k2.6` non-thinking	fixed 0.6
`kimi-k2.5` thinking	fixed 1.0
`kimi-k2.5` non-thinking	fixed 0.6

Passing any other value returns HTTP 400. Moonshot has already tuned each model to its optimal temperature, so this MCP deliberately omits the parameter and lets the API use its built-in default. Use the thinking flag to switch between the 1.0 / 0.6 modes for K2.6 / K2.5.

Why Kimi for vision?

Longest context: kimi-k3 ships with 1M tokens — useful for analyzing long documents alongside images.
Strong Chinese & English OCR.
Native video understanding on kimi-k3 / kimi-k2.7-code / kimi-k2.6 (not exposed by this MCP yet — file upload only).
Cost-effective: ¥2/M output for the flagship.