vision-mcp
MCP server that gives Claude and other clients vision capabilities including screenshot-to-code, OCR, error diagnosis, diagram analysis, data-viz insights, UI diff checks, and video analysis via any OpenAI-compatible vision endpoint.
README
vision-mcp
An MCP (Model Context Protocol) server that gives Claude and other MCP clients vision: screenshot → UI code, OCR, error-diagnosis, diagram analysis, data-viz insights, UI diff checks, and video analysis.
Vendored from @z_ai/mcp-server by Z.AI (Apache-2.0) and extended to be provider-agnostic — point it at any OpenAI-compatible vision endpoint with base URL + model + API key.
Status
Working. All 8 tools verified booting over MCP stdio (see build/).
Version 0.1.4 (upstream).
What it does
Eight tools cover image and video analysis:
| Tool | What it does |
|---|---|
ui_to_artifact |
Convert UI screenshots into frontend code, AI prompts, design specs |
extract_text_from_screenshot |
OCR: code, terminal output, logs, docs from screenshots |
diagnose_error_screenshot |
Analyze error messages/stack traces, suggest fixes |
understand_technical_diagram |
Explain architecture, flowchart, UML, ER diagrams |
analyze_data_visualization |
Insights from charts and graphs |
ui_diff_check |
Visual regression: expected vs actual UI |
analyze_image |
General-purpose image analysis (fallback) |
analyze_video |
Video content analysis (URL or local file; local file ≤8MB, ZAI/Zhipu only) |
Install
Requires Node.js 18+.
Easiest — no clone, no npm publish. npx runs the server straight
from the GitHub repo (cached after the first run):
npx -y github:nikkoxgonzales/vision-mcp
Or clone and install locally:
git clone https://github.com/nikkoxgonzales/vision-mcp.git
cd vision-mcp
npm install
Usage
Three environment variables configure the server. AI_API_KEY is
required; AI_BASE_URL and AI_VISION_MODEL select the provider and
model.
| Variable | Required | Description |
|---|---|---|
AI_API_KEY |
Yes | API key for the provider |
AI_BASE_URL |
No | OpenAI-compatible base URL (e.g. https://api.openai.com/v1/). Default: Zhipu or Z.AI preset via AI_MODE. |
AI_VISION_MODEL |
No | Vision model name (default: glm-4.6v) |
Convenience presets via AI_MODE (or PLATFORM_MODE) fill in
AI_BASE_URL when it isn't set: ZAI → https://api.z.ai/api/paas/v4/,
ZHIPU (default) → https://open.bigmodel.cn/api/paas/v4/. An explicit
AI_BASE_URL always wins.
Legacy aliases
Z_AI_API_KEY,Z_AI_BASE_URL,Z_AI_VISION_MODEL,Z_AI_MODE,ZAI_API_KEYandZ_AI_*tunables are still accepted;AI_*takes precedence when both are set.
Size limits: local images ≤5MB, local videos ≤8MB (URLs are passed
through unchecked). AI_VISION_MODEL_MAX_TOKENS defaults to omitted —
the provider's own output cap applies; set it explicitly to cap output.
AI_RETRY_COUNT (default 1) controls API retries; only transient
failures (network, timeout, 5xx) are retried.
Claude Code
# Z.AI
claude mcp add vision-mcp --env AI_API_KEY=your_key AI_MODE=ZAI -- npx -y github:nikkoxgonzales/vision-mcp
# Zhipu
claude mcp add vision-mcp --env AI_API_KEY=your_key AI_MODE=ZHIPU -- npx -y github:nikkoxgonzales/vision-mcp
# Any OpenAI-compatible provider
claude mcp add vision-mcp \
--env AI_API_KEY=your_key \
--env AI_BASE_URL=https://api.openai.com/v1/ \
--env AI_VISION_MODEL=gpt-4o \
-- npx -y github:nikkoxgonzales/vision-mcp
With a local clone, swap npx -y github:nikkoxgonzales/vision-mcp
for node path/to/vision-mcp/build/index.js.
Other MCP clients — .mcp.json
Clients that read a project-level .mcp.json (VS Code, Cursor, Windsurf,
...): drop this at the project root.
{
"mcpServers": {
"vision-mcp": {
"command": "npx",
"args": ["-y", "github:nikkoxgonzales/vision-mcp"],
"env": {
"AI_API_KEY": "your_api_key",
"AI_MODE": "ZAI"
}
}
}
}
For any other OpenAI-compatible provider, set the base URL and model
instead of AI_MODE:
{
"mcpServers": {
"vision-mcp": {
"command": "npx",
"args": ["-y", "github:nikkoxgonzales/vision-mcp"],
"env": {
"AI_API_KEY": "your_api_key",
"AI_BASE_URL": "https://api.openai.com/v1/",
"AI_VISION_MODEL": "gpt-4o"
}
}
}
}
With a local clone, swap npx/args for command: "node",
args: ["path/to/vision-mcp/build/index.js"].
Upstream platform docs: Z.AI Vision MCP · Zhipu Vision MCP
Project layout
build/— compiled server (vendored from@z_ai/mcp-server0.1.4)package.json— package metadata; deps:@modelcontextprotocol/sdk,zod
Changelog
2026-08-12— hardening pass: API-key validation now matches real placeholders (real keys containing "api"/"key" no longer rejected); whitespace-only keys rejected; undocumentedANTHROPIC_AUTH_TOKENfallback removed;AI_MODEmatching case-insensitive; invalid numeric env values fall back to defaults;max_tokensomitted when unset; retries limited to transient failures andAI_RETRY_COUNTnow wired;analyze_videorestricted to ZAI/Zhipu (generic endpoints don't acceptvideo_url); server identityvision-mcpv0.1.4.2026-08-12— v0.1.4 vendored; provider-agnostic config (AI_BASE_URL+AI_VISION_MODEL+AI_API_KEY, legacyZ_AI_*aliases accepted); explicit base URL wins overAI_MODE; provider-specific request params omitted for custom endpoints.
Credits
- Z.AI — upstream
@z_ai/mcp-server(Apache-2.0) - Contributors: Chao Gong, Lei Yuan
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。