Server For CAP

Build a Production-Grade MCP Server for CAP

Overview

The server exposes a set of MCP tools and resources that let Claude interact with a document processing backend. It uses the MCP SDK with a StreamableHTTPServerTransport, meaning each session is stateful and tracked by a mcp-session-id header.

Authentication

The token is never exposed to the client — it lives only in the server process via process.env.BEARER_TOKEN. The MCP endpoint itself (POST /mcp) does not require authentication from Claude's side

Tool Reference

list_documents : Lists all documents, with an optional status filter.

get_document : Retrieves full details for a single document, including its parsed invoice and line items.

upload_document : Uploads a document as a base64-encoded string to the backend.

process_document : process_document

get_audit_log : Returns all audit log entries for a given document.

Resources

documents://pending : List of documents with PENDING status

documents://stats : Returns aggregate statistics across all documents:

    Total document count
    Count of DONE documents
    Count of FAILED documents
    Average AI confidence score (from expanded invoice data)

Retry Strategy

All OData calls go through withRetry(), an exponential backoff wrapper:
Attempt 1 ──► fail ──► wait 500ms
Attempt 2 ──► fail ──► wait 1000ms
Attempt 3 ──► fail ──► throw

Connecting from SAP BTP / Cloud Foundry

Create an Node.js application. npm install .env.example .env # fill in APP_URL and BEARER_TOKEN node index.js # To run the application locally

Push to CF

cf push # To push the code to CF After deploying the application i tested from the POSTMAN.

NOTE : By using the VScode, I cloned the application and i tested via the official MCP Inspector. node index.js # To run the application locally. For MCP Inspector # npx @modelcontextprotocol/inspector.

NOTE : When iam trying to connect the Claude code with my application, It is not working. (It may work with a licensed Claude account).