Build a Local MCP Server for a Cancer Detection Inference API (Research Prototype)

We’re researching AI-assisted cancer detection for hospital pathology workflows. We have a FastAPI inference service that classifies tissue patches from the public PatchCamelyon (PCam) dataset as cancerous or healthy, with endpoints for predictions, Grad-CAM heatmaps, and model info. We’d like AI assistants such as Claude Desktop and Cursor to use this model directly. You’ll build a Model Context Protocol (MCP) server that wraps our existing API and exposes it as clear, well-described tools. This is a research prototype for internal evaluation. All testing uses public PCam images. Scope: MCP tools: classify_tissue_patch: accepts an image (file path or base64) and returns label, confidence, and model version explain_prediction: returns the Grad-CAM heatmap as an image the client can display, with a one-line summary get_model_info: returns model name, version, input size, and decision threshold Backend and integration: Calls our inference API over HTTP, with base URL and API key set via environment variables stdio transport for local MCP clients Input validation, timeouts, and clear error messages returned to the client Every tool response includes a short research-use disclaimer Deliverables: Source code in a private GitHub repo (Python with FastMCP or the official MCP SDK, or TypeScript with the official MCP SDK) Automated tests for each tool, invalid inputs, and API timeouts, using a mocked inference API Verified working connection in Claude Desktop and the MCP Inspector, with config snippets in the README README covering setup, environment variables, and how to add new tools A short screen recording showing an AI assistant classifying an image and displaying its heatmap Requirements: Hands-on experience building MCP servers Strong Python or TypeScript backend skills Solid API integration fundamentals: validation, timeouts, error handling Screening questions: Share a link to or describe an MCP server you’ve built, and which clients you tested it with. How would you write tool descriptions so an AI assistant picks the right tool and passes valid inputs? How would you return a heatmap image through MCP so it displays well in Claude Desktop? How we’ll work: We’ll share access to the inference API and its OpenAPI docs once the contract starts. Share the repo with classify_tissue_patch working in the MCP Inspector within 48 hours, then complete the remaining tools, tests, and documentation.