The scope is deliberately narrow — no window targeting, no app list, just the full screen and a question — which makes it the shortest path from "look at this" to an answer. The design consequence worth knowing is that analysis runs against your own Anthropic key rather than the client's session, so every capture is a billable vision call you control and pay for. Pinning the version in the config is not fussiness; the README calls it out because npx caching has bitten people here.
A screenshot-and-analyse tool for MCP clients. It captures the whole screen and passes the image to the Claude Vision API along with a prompt you supply, so the assistant reasons about what is actually rendered rather than about your description of it.
- Capture and analysis as a single call, taking a `prompt` for the question and a `modelName` to choose which model looks at it — `captureAndAnalyzeScreen`
- A trivial echo tool for checking the connection is alive before you debug anything else — `helloWorld`
- Both stdio and SSE transports, since clients differ — Claude Desktop uses stdio, while Cursor, Cline and Windsurf support either
- A Smithery deployment path with a Dockerfile and config in the repository, for hosting it over WebSocket instead of running it locally
An Anthropic API key, supplied as `ANTHROPIC_API_KEY` in your client config or as a command-line argument — the vision call is billed to you, not to the client. Node.js 18+. Install globally with npm as `screen-view-mcp`, or launch it with npx; the README recommends pinning the exact version (2.0.15 in package.json) in your config to avoid npx cache surprises. For SSE, the built entry takes a port and host alongside the key.
One command plus a key — npm install -g screen-view-mcp, then supply credentials