eyes-mcp
eyes-mcp
让任何纯文本 LLM 拥有眼睛。为你的编程代理提供本地视觉。
一条命令 · 零 API 密钥 · 一切都不离开你的机器
DeepSeek、GLM、Qwen-Coder、Llama…优秀的模型,却全都看不见。
❌ 没有眼睛
你把截图粘贴给代理(通过 Claude Code / Codex / Cursor 运行纯文本模型):
> Here's the error in my UI, fix it [screenshot.png]
I'm sorry — I cannot see images. Please describe the error in text.Related MCP server: OpenSight MCP
✅ 有眼睛
代理改为调用本地 VLM + OCR,自己读取截图:
> Here's the error in my UI, fix it [screenshot.png]
I see a React hydration error in `CartDrawer.tsx:142`. The OCR shows:
"Hydration failed because the server rendered HTML didn't match the client." …快速开始
git clone https://github.com/JamesbbBriz/eyes-mcp
cd eyes-mcp && ./scripts/install.sh就这样。安装程序会:
询问你想使用哪个模型,并根据你的内存和 GPU 给出推荐(使用
EYES_PRESET或--yes可跳过该问题);安装依赖并下载模型(约 0.3 至 3.5GB,支持断点续传);
通过读取你的 Claude Code / Codex / Cursor 配置,并对照模态数据库检查每个模型,来检测哪些代理运行的是纯文本模型;
只在需要的地方注册 eyes-mcp。多模态代理会自动跳过。
# options:
EYES_PRESET=fast ./scripts/install.sh # Qwen3.5-0.8B, natively multimodal
HF_ENDPOINT=https://hf-mirror.com ./install.sh # mainland-CN mirror
./install.sh --yes # accept all recommendations, no prompts
./install.sh --dry-run # preview without changing anything只想做模态检查?运行 python3 scripts/detect_modality.py
要求:Python ≥3.11、llama.cpp(brew install llama.cpp),约 1GB 内存。
手动注册
跳过了自动安装,或者遇到安装程序不认识的代理?手动添加即可。
Claude Code(~/.claude.json → mcpServers):
"eyes-mcp": {
"command": "uv",
"args": ["--directory", "/ABS/PATH/eyes-mcp", "run", "eyes-mcp"],
"env": { "EYES_PRESET": "lfm-450m" }
}Codex(~/.codex/config.toml):
[mcp_servers.eyes-mcp]
command = "uv"
args = ["--directory", "/ABS/PATH/eyes-mcp", "run", "eyes-mcp"]
env = { EYES_PRESET = "lfm-450m" }Cursor(.cursor/mcp.json):与 Claude Code 的结构相同。
重启代理,然后问:“这张截图里有什么?”
工具
工具 | 引擎 | 用途 |
| 通过 llama.cpp 运行的 VLM | 描述、UI 理解、视觉问答 |
| RapidOCR (onnx) | 密集文本:终端、文档、表格;快速且精准 |
模型预设
预设 | 模型 | 下载 | 内存 | 许可证 | 备注 |
| SmolVLM2-256M | ~0.3GB | ~1GB | Apache-2.0 | 最小的实用 VLM |
| LFM2.5-VL-450M | ~0.4GB | ~1.2GB | 经过测试;启动最快 | |
| Qwen3.5-0.8B | ~0.7GB | ~1.8GB | Apache-2.0 | 原生多模态(图像 + 视频) |
| GLM-OCR | ~1.4GB | ~3.5GB | MIT | 密集文本 / 文档冠军(每月 300 万+ 下载) |
| Qwen3.5-2B | ~2GB | ~3.5GB | Apache-2.0 | 最佳质量/体积平衡 |
| Qwen3.5-4B | ~3GB | ~6GB | Apache-2.0 | 最高档(建议 GPU) |
隐藏附加项(同样只需一条命令):smol500(SmolVLM2-500M)、paddle(PaddleOCR-VL-1.6)、qwen3-2b(Qwen3-VL-2B)。
其他任何 GGUF 也都可以。将环境变量指向它,完全跳过预设:
EYES_MODEL_DIR=~/models/my-vlm VLM_MODEL_FILE=model-Q4.gguf VLM_MMPROJ_FILE=mmproj.gguf未内置为预设的良好候选项:LFM2.5-VL-1.6B/3B、InternVL3.5-2B/4B、MiniCPM-V-4.6、DeepSeek-OCR、dots.ocr、gemma-3n-E2B、moondream2。任何 llama.cpp 支持且带有 mmproj 文件的模型都可以。
随时切换:设置 EYES_PRESET,再次运行 ./scripts/download_models.sh。不确定选哪个?python3 scripts/choose_model.py 会显示你的内存/GPU 并标出推荐项。
工作原理
Claude Code / Codex / Cursor
│ MCP stdio
▼
eyes-mcp (stateless, mcp SDK 2.x)
├─ analyze_image → llama.cpp llama-server (local VLM) "understand"
└─ ocr_image → RapidOCR (onnx, ~20MB) "extract text"生命周期随代理启动/结束:VLM 服务器在 MCP 启动时生成,在代理退出时关闭,因此你永远不会留下孤儿进程,也无需守护进程。
动态端口:VLM 从不绑定固定端口(告别“8080 已被占用”),因此它可与你本地的其他服务共存。
复用外部 VLM:如果你已经在
VLM_BASE_URL运行了一个 VLM,eyes-mcp 会直接使用它,而不是另起一个。
为什么
当前最便宜也最好的编程模型(DeepSeek-V4-Flash、GLM-5.x、Qwen-Coder)都是纯文本的。每个工具链都默认你能粘贴截图,而这些模型无一例外都会在这上面悄然失败。eyes-mcp 就是那块缺失的伴生组件:一个本地小 VLM 加上 OCR,并以你的代理已经理解的生命周期方式包装起来。
路线图
延迟启动 VLM(在首次工具调用时生成,而非 MCP 启动时)
screenshot_analyze(直接截取屏幕,无需文件)PDF 页面 → 视觉
npx eyes-mcp一行命令安装程序针对不同模型的提示词模板(llama.cpp OCR 模型需要特定的提示词)
常见问题
我的代理模型有关系吗? 唯一需要注意的是:它必须是纯文本模型,这个工具才有用。多模态模型(GPT、Claude、GLM-V)本身就能看图,所以不必在意。
需要 GPU 吗? 不需要。它在 CPU 上运行良好,而且 llama.cpp 会自动利用 Apple Metal 或 CUDA(如果存在)。
模型存储在哪里? ~/.eyes-mcp/models/<preset>/。删除它们即可重置。
许可证
MIT。模型权重保留各自的许可证(见预设表);它们在安装时下载,绝不在此重新分发。
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Related MCP Connectors
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Give your agent eyes: screenshots of any web page it can see, pages as markdown, and PDFs.
Web scraping, browser automation, screenshots and AI data extraction for AI agents via MCP.
SnapRender gives AI agents eyes on the web through a hosted MCP server. Send a public URL and get the page back as an image, a PDF or clean markdown, with no browser to run or maintain. Capture the viewport or the full page as PNG, JPEG, WebP or PDF, at desktop size or with a phone, tablet or laptop preset, in light or dark mode. Ads and cookie banners are removed by default, and you can hide elements or click one before the capture. extract_content reads the same page as markdown, text, article, links or metadata, so an agent can see a page and read it in one session. Made for monitoring and client work: batch up to 50 URLs in one job, check the cache for free before spending a capture, share signed links that render without an API key, and get webhooks when a batch finishes or a scheduled capture spots a change. Track competitors' pricing and landing pages, or send clients current screenshots of their sites. Connect with OAuth (Google, GitHub or email) or an API key. Works with Claude, ChatGPT, Cursor, VS Code and any MCP client. Free plan: 200 screenshots a month, no credit card. Paid plans from $9 a month for 2,000 up to $199 a month for 200,000. Home: https://snap-render.com Docs: https://snap-render.com/ai-agents
Related MCP Servers
- AlicenseAqualityDmaintenanceBridges vision models to text-only coding models using Florence-2, enabling non-vision LLMs to describe images, extract text, and analyze screenshots via MCP tools.6MIT
- AlicenseNot gradedqualityDmaintenanceMulti-backend AI vision for MCP agents. Analyze images, screenshots, and documents using local Ollama models or cloud APIs like OpenAI, Google Gemini, and OpenRouter.MIT
- AlicenseNot gradedqualityCmaintenanceLocal vision-capable MCP server that lets AI agents describe screenshots, UI, charts, and photos via vision and OCR tools, with support for multiple providers and automatic fallback.5MIT
- AlicenseNot gradedqualityNot gradedmaintenanceAdds image recognition and UI grounding capabilities to text-only LLMs through MCP tools, supporting local and cloud vision backends.MIT