A document comparison rule is only enforced if every pair gets compared, not the ones someone had time for. With the GroupDocs.Comparison.Mcp server, one prompt makes Claude Desktop, Claude Code, Cursor or GitHub Copilot compare every version pair in a folder locally and report which documents changed:

For every -v1/-v2 pair in my documents folder, compare the two and tell me which invoices changed and how. One line per invoice.

The step-by-step version with config and troubleshooting is in the documentation: How to batch-compare document versions in a folder.

Which checks fail when comparison is manual?

Three situations come up repeatedly, and in each of them the unchecked document is the one that matters:

  • A re-issued batch. You need to confirm that the new batch matches what was approved.
  • A vendor’s “unchanged” statement. The vendor says the documents did not change; you want evidence.
  • A migration. Documents moved between systems or formats, and the output must equal the input.

In each case the cost is not the comparison but the discipline of running it for every file. An agent that loops over a folder removes the discipline problem.

How does the workflow run?

  1. Name the pairs consistently. Any convention works, for example invoice-001-v1.pdf and invoice-001-v2.pdf, in one folder.
  2. Point the server at the folder. Set GROUPDOCS_MCP_STORAGE_PATH to it. Files are resolved by name, so the agent passes invoice-001-v1.pdf, not a full path.
  3. Check the license mode. Ask: “What is the license status of the comparison server?” The agent calls get_license_status.
  4. Run the sweep. Use the prompt from the start of this post. The agent calls analyze_changes once per pair.
  5. Render only what matters. Ask: “Now produce marked-up copies for the three that changed.” The agent calls compare for those three, and each result is saved as <source-stem>_compared<ext>. Existing files are never overwritten.

Use analyze_changes for the sweep and compare only for the exceptions. Rendering thirty marked-up documents that nobody opens is thirty times the work for the same answer.

What does the agent decide and what does the engine do?

Step Agent Engine
Find the -v1/-v2 pairs Matches the names
Compare each pair Calls analyze_changes Detects changes, per page and table cell
Decide a pair is unchanged Reads No changes detected Returns that result
Report to you Writes one line per pair
Produce files for the changed ones Calls compare Renders the marked-up document

The pass/fail judgment is made by the engine’s result, not by the agent’s impression of the text.

Example session

Abridged and illustrative; the real wording depends on your client and model.

You:    For every -v1/-v2 pair in my documents folder, compare the
        two and tell me which invoices changed and how. One line
        per invoice.

Agent:  [calls analyze_changes: invoice-001-v1.pdf, invoice-001-v2.pdf]
        [calls analyze_changes: invoice-002-v1.pdf, invoice-002-v2.pdf]
        [calls analyze_changes: invoice-003-v1.pdf, invoice-003-v2.pdf]

        invoice-001: no changes detected
        invoice-002: changes detected (see the list for the pages)
        invoice-003: changes detected (see the list for the pages)

What are the limits of a batch run?

  • Evaluation cap, per document. Without a license, only the first 2 pages of each document are compared and results carry a watermark, so a batch report is quietly incomplete. Check get_license_status before you trust a sweep.
  • Same format within a pair. Convert first if the two files differ in format.
  • A wrong name is not fatal. The tool answers with the list of files it can see, and the agent corrects itself and retries.
  • Long sweeps take time. Each pair is a real comparison; a folder of large PDFs takes minutes, not seconds. Ask for progress as it goes.

FAQ

Can I compare multiple PDF files at once with AI? Yes. Put the pairs in one folder and ask the agent to compare each pair. It calls analyze_changes once per pair and gives you one consolidated answer.

How do I detect unauthorized changes in documents? Compare the approved version with the version in circulation and ask the agent to list the pairs where changes are detected. Every pair that returns No changes detected has the same content, provided the server is licensed: under evaluation mode only the first 2 pages of each document are compared.

Do the files leave my machine? No. The comparison runs in a local process against the folder you configure. The text of your prompt and the lines the agent reports go to the model your client uses.

Go deeper