Vision MCP Server

A remote MCP (Model Context Protocol) server exposing 7 vision tools over the 讯飞 MaaS Anthropic-compatible model API (xopkimik26). Supports Claude Code, Codex, Cline, and any MCP-compatible client via Streamable HTTP.

Features

• 7 specialized vision tools (UI→code, OCR, error diagnosis, diagram understanding, data-viz analysis, UI diff, general image analysis)

• Multi-API-key round-robin with cooldown + retry (401/403/429/5xx/network)

• Global (100) and per-key (20) concurrency control

• Image preprocessing: PNG/JPEG/GIF/WebP/BMP → JPEG Q80 with resolution caps (2048 standard / 4096 OCR / 1536 diff)

• /health status endpoint

Related MCP server: read-image-mcp

Quick start

npm install
npm run build
API_KEYS=key1,key2 npm start

Server listens on PORT (default 3000). Endpoints: POST /mcp (MCP), GET /health.

Docker

echo "API_KEYS=key1,key2" > .env
docker compose up -d

Configuration (env)

Connect a client

Claude Code:

claude mcp add -s user vision-mcp-server --transport http http://localhost:3000/mcp

Manual config (.claude.json):

{
  "mcpServers": {
    "vision-mcp-server": { "type": "http", "url": "http://localhost:3000/mcp" }
  }
}

Development

npm test          # node:test over src/**/*.test.ts
npm run typecheck # tsc --noEmit
npm run dev       # tsx watch src/index.ts

Notes

• The model API accepts base64 images only (URLs return 400). Clients (e.g. Claude Code) auto-encode local files.

• Token cost scales with resolution, not file size — hence JPEG Q80 + resolution caps.

• Transport is stateless (sessionIdGenerator: undefined) for horizontal scalability.