Two tools, one for a URL and one for a path, is the whole surface — which is right when the job is simply "what is in this picture". The detail that causes most failures is path translation: the server resolves the path in its own environment and does not rewrite it, so a Windows path handed to a WSL server will not resolve.
A small TypeScript server that sends an image to the GPT-4o-mini model and returns an analysis of what is in it. It accepts either a URL, which it validates first, or a local file, which it reads and Base64-encodes before sending.
- `analyze_image` takes an image URL and describes its content
- `analyze_image_from_path` does the same for a local file
- URL validity is checked before the call, so a broken link fails early rather than as a model error
An OpenAI API key in `OPENAI_API_KEY`. Install via Smithery, or clone and run `npm install` then `npm run build` and point your client at `dist/index.js` with `node`. Paths are interpreted in the environment where the server runs — if the server is under WSL and your client has a Windows path, the client must convert it before calling. A TS7016 type error about `mime-types` during the build does not affect execution.
One command plus a key — npx -y @smithery/cli install @champierre/image-mcp-server --client claude, then supply credentials