Text Agent (e.g. DeepSeek / Claude)
  → MCP Client → inspect_image tool
  → Local image sandbox + Sharp crop/scale
  → MiMo-V2.5 (vision model, Anthropic Messages API)
  → VisualObservation JSON
  → Agent continues reasoning

git clone https://github.com/wenren93/mimo-vision-mcp.git
cd mimo-vision-mcp
npm install
cp .env.example .env
# Edit .env and set MIMO_API_KEY
npm run build
npm run demo -- /path/to/image.png "What is in this image?"

npm run start              # STDIO mode
npm run inspect            # MCP Inspector UI

claude mcp add vision \
  -e MIMO_API_KEY='your-key' \
  -e MIMO_BASE_URL='https://api.xiaomimimo.com/anthropic' \
  -e VISION_MODEL='mimo-v2.5' \
  -e VISION_ASSET_ROOT='/absolute/path/to/assets' \
  -- node /absolute/path/to/dist/server.js

DeepSeek V4 Pro（纯文本主 Agent，Anthropic Messages 兼容 API）
  -> MCP Client
  -> 本地 inspect_image 工具
  -> 本地图片沙箱 + Sharp 裁剪/缩放
  -> MiMo-V2.5（Anthropic Messages 兼容 API）
  -> VisualObservation JSON
  -> DeepSeek 继续推理或操作浏览器

cd mimo-vision-mcp
npm install
cp .env.example .env

# MCP 视觉工具需要
MIMO_API_KEY=你的小米MiMo密钥
MIMO_BASE_URL=https://api.xiaomimimo.com/anthropic
VISION_MODEL=mimo-v2.5

# 仅端到端 Demo Agent 需要
DEEPSEEK_API_KEY=你的DeepSeek密钥
DEEPSEEK_BASE_URL=https://api.deepseek.com/anthropic
TEXT_MODEL=deepseek-v4-pro

VISION_ASSET_ROOT=./assets
VISION_MAX_FILE_MB=10
VISION_MAX_PIXELS=40000000
VISION_TIMEOUT_MS=45000

npm run demo -- /绝对路径/page.png "图中登录按钮在哪里？"

npm run build
npm run start

npm run inspect

npm run add-asset -- /绝对路径/page.png

claude mcp add vision \
  -e MIMO_API_KEY='你的 MiMo API Key' \
  -e MIMO_BASE_URL='https://api.xiaomimimo.com/anthropic' \
  -e VISION_MODEL='mimo-v2.5' \
  -e VISION_ASSET_ROOT='/绝对路径/mimo-vision-mcp/assets' \
  -- node /绝对路径/mimo-vision-mcp/dist/server.js

import_image({ sourcePath: "/你的路径/page.png" })
→ { assetId: "img_....png" }
→ inspect_image({ assetId: "img_....png", goal: "找到登录按钮" })

{
  "assetId": "img_生成的ID.png",
  "goal": "找到登录按钮并返回位置",
  "mode": "ui",
  "resolution": "auto"
}

{
  "mcpServers": {
    "vision": {
      "command": "node",
      "args": ["/绝对路径/mimo-vision-mcp/dist/server.js"],
      "env": {
        "MIMO_API_KEY": "你的密钥",
        "MIMO_BASE_URL": "https://api.xiaomimimo.com/anthropic",
        "VISION_MODEL": "mimo-v2.5",
        "VISION_ASSET_ROOT": "/绝对路径/mimo-vision-mcp/assets"
      }
    }
  }
}

MIMO_BASE_URL=http://127.0.0.1:8000
MIMO_API_KEY=local
VISION_MODEL=mimo-v2.5

vllm serve XiaomiMiMo/MiMo-V2.5 \
  --served-model-name mimo-v2.5 \
  --host 127.0.0.1 \
  --port 8000 \
  --trust-remote-code \
  --reasoning-parser qwen3

const assetId = `shot_${crypto.randomUUID()}.png`;
await page.screenshot({
  path: path.join(process.env.VISION_ASSET_ROOT!, assetId),
  fullPage: false,
});

const clickX = (box.x + box.width / 2) * viewport.width;
const clickY = (box.y + box.height / 2) * viewport.height;
await page.mouse.click(clickX, clickY);

# 安装依赖
npm install

# 开发模式运行
npm run dev

# 类型检查
npm run typecheck

# 代码检查
npm run lint
npm run lint:fix

# 代码格式化
npm run format
npm run format:check

# 运行测试
npm test