How to run a document review cycle with an AI agent Leave feedback

A review round is rarely “add a comment”. It is: what is still open, who raised it, answer it, resolve it, and produce a clean copy for the next audience. All five are tool calls.

Note

The commands and config snippets on this page are for the .NET build of the server — the only platform available today. Installation and client setup: MCP server for .NET. Other platforms will expose the same tools with their own launch command; everything else on this page applies unchanged.

List the annotations in contract_annotated.pdf. Which ones have no replies?

get_annotations returns every annotation with its id, author, page, message, and replies. “Unanswered” is just an empty replies array — the agent filters the list it already has, without re-reading the document.

Reply to Maria’s comment that finance approved the terms on 12 March. Sign it as Alex.

add_reply takes the annotationId from step 1, the comment, and a userName. Supply the name: a thread without authors cannot be filtered later, and remove_replies can clean up by author only if the author was recorded.

There is no “resolved” flag in the model — resolution is a convention you choose:

• Reply and leave it — the thread is the audit trail; nothing is deleted.
• Remove the annotation — remove_annotations by id once an issue is closed, keeping only open points.
• Edit the message — update_annotation to prefix [RESOLVED], which keeps the history visible in any viewer.

Pick one and tell the agent which: “mark resolved by prefixing the message, do not delete anything” is a perfectly good instruction.

The document changes; the comments should not be lost:

Export the annotations from the reviewed copy, then import them into contract-v3.pdf.

export_annotations writes an XML file; import_annotations merges it into the new draft. Coordinates carry over as-is, so on a re-flowed document a note may land near rather than exactly on its original text — worth a preview pass afterwards.

The same pair merges two reviewers’ copies into one.

Remove all annotations and give me a clean copy for the counterparty.

remove_annotations with no ids clears everything into a new file. Ask the agent to confirm the count before it runs — this is the one call in the workflow with no undo.

Two habits make this workflow trustworthy:

• Always name the file you mean. Each step writes a new file; the agent should pass the previous result forward, not the original.
• Check the licence before a formal review. In evaluation mode every page carries a trial badge, which is not something to send to a counterparty — get_license_status answers in one call.