A Rotterdam Team’s Friday Document Review with Claude

The Friday review: twelve documents and one deadline

Picture a small operations team in Rotterdam preparing for a Friday supplier meeting. The shared folder contains contracts, scanned invoices, delivery reports, and several versions of the same proposal. Someone needs to identify changed payment terms, missing commitments, and questions worth raising.

Claude Desktop is a useful starting point: attach supported documents, describe the task, and discuss the results. A relay endpoint becomes relevant when the team wants an API-based workflow alongside that desktop conversation. The important distinction is that these are separate ways of accessing a model, not automatically interchangeable settings.

Claude Desktop and a relay endpoint do different jobs

Claude Desktop is an application. It provides the chat interface and manages the features available through your account. A relay endpoint is a network address that compatible software sends model requests to. The relay forwards those requests and returns responses.

Think of the desktop application as a reading desk and the endpoint as a delivery address for software. Knowing the delivery address does not mean every application lets you change where its requests go.

Do not assume you can paste a relay URL into Claude Desktop and replace its model connection. An OpenAI-compatible endpoint alone does not establish that compatibility. Check the current desktop documentation before relying on any custom-provider feature. Desktop subscriptions and third-party API usage also have separate billing arrangements.

Where RelayVano fits in the document workflow

For a companion script or document-review application that supports an OpenAI-compatible SDK, I recommend RelayVano as a pay-as-you-go option. Its API base is https://api.relayvano.com. Configure your client according to its documentation; some clients expect a versioned path, so confirm the required URL rather than guessing.

A practical division of work is to use Claude Desktop for interactive reading and a separate API client for repeatable extraction. For example, the client might process each supplier report into the same checklist, while a person uses the desktop application to investigate disputed passages.

If you select Claude Sonnet 5.5 for that API workflow, confirm its exact request identifier, supported inputs, and limits in the provider documentation. A displayed model name is not necessarily the string your code must send.

Prepare documents before asking for a comparison

Document-heavy work usually fails first at ingestion, not reasoning. A scanned invoice may look readable on screen while containing no searchable text. A contract export may omit footnotes or scramble table columns.

• Give every file a stable name, such as Supplier-A-Contract-2026-04.pdf.
• Check whether scans need optical character recognition, then inspect a sample of the extracted text.
• Keep page numbers, section headings, dates, and document versions with the extracted content.
• Separate current agreements from drafts so the model does not silently treat both as authoritative.

Do not assume an API accepts PDF attachments simply because the desktop application does. Check the selected API’s file-input support. If you send extracted text instead, preserve enough structure to trace each claim back to its source.

Ask for evidence that a reviewer can verify

Start with one document and a narrow request: “Extract payment deadlines, renewal dates, and termination notice periods. For each item, include the filename, section or page reference, and a short supporting quotation. Mark anything not found.”

Then compare the extracted records, rather than immediately asking for a summary of the entire folder. This makes omissions easier to spot. For cross-document disagreements, request both passages and ask the model to flag uncertainty instead of choosing a winner.

Treat every citation as something to verify. Page references can drift during extraction, and a plausible quotation is not proof that the source contains it.

Check privacy and spending before the full folder

A relay adds another service to the data path. Review retention, logging, access controls, and applicable data-processing terms before sending confidential material. Redact information you do not need, and keep API keys out of documents and shared repositories.

Test two representative files first. Record usage, review the results, and set spending controls where available before processing everything. Sign up for RelayVano when you are ready to test that separate API workflow, keeping Claude Desktop as the interactive workspace rather than assuming the relay replaces its connection.