The design pressure moved from 150 to 175 psig. Fourteen datasheets reference it, plus the line list and two sections of the narrative spec. You point ChatGPT Work at the project folder, ask it to update them, and five minutes later it reports back that it did the job.

Do you trust it? It edited fourteen files in a folder that the rest of your team works out of. Maybe it caught the pressure in the notes block on sheet 3. Maybe it also reworded a materials call-out nobody asked it to touch. From a summary and a folder of modified files, you can't tell those two cases apart without reading all fourteen.

If the agent changed fourteen datasheets, you see all fourteen. And if it actually made fifteen edits, you see the fifteenth as well.

The copy-paste era had one accidental safety feature

Until this year, using AI on engineering documents meant a chat window. Paste in a spec section, get a rewrite, paste it back. This was slow, frustrating, and mistake-prone, but nothing changed inside a file without a human watching that change.

Agents that work on real files remove both the tedium and the oversight. You send a prompt and get back a folder of modified files and a paragraph about them. You don't know whether the changes are correct. And SharePoint's version history only tells you that a file changed, not what changed inside it.

Point the agent at Kord and the edits come back as a diff

Kord runs a Model Context Protocol (MCP) server. Claude Code, Claude Cowork, ChatGPT Work, and Cursor can all connect with one URL and then work with your document sets. Your file storage doesn't move - Kord tracks the files that live in your shared drive - so the agent edits them in place. Kord independently tracks every file that changes, and shows you exactly what those changes are.

Connecting is easy - one URL plus a browser sign-in with the Google or Microsoft account you already use. The agent sees the teams, repositories, and review sessions your account sees, and nothing else. ChatGPT and Claude both let you control which actions a connector can take on its own, and you can revoke Kord's access from either side at any time.

When it finishes, the agent opens a review session for the files it changed. The review session compares the new version to the last published version: line by line on text, cell by cell on spreadsheets, with changed text called out on PDFs and Word documents.

• You see every change, not just what the agent reports. Edits in files you never asked about show up alongside the ones you expected.
• No opening two revisions side by side to hunt for three changed lines. The changed lines are highlighted for you.
• Assign required reviewers, make comments and markups, then fix it yourself or simply ask the agent.
• Every change is versioned, with an author and a timestamp.

It reviews as well as it edits

The connection runs both directions. An agent can read a document set the way a reviewer does: pull the text out of a file, walk its revision history, look at what is already under review and what changed in it. Then it can act on what it finds instead of handing you a list to act on.

• Ask it to read the spec against the datasheets and flag where they disagree. It comments on the line or cell, inside the review session, where whoever fixes it is already looking.
• Something worth tracking but not worth fixing in this pass becomes an issue on the repository, linked to the files it affects.
• Its comments are ordinary comments: threaded, resolvable, and written into a drive-linked Word file as a native Word comment.
• It can answer the tedious question quickly: which review sessions are waiting on your approval, across every repository you have access to.

Sign-off stays with a person

An agent connected to Kord cannot approve. It cannot mark a review as passed, request changes as a reviewer, or publish a revision. Those are left out on purpose.

You set the bar per review session. Reviewers marked required block publishing until they approve; additional reviewers can weigh in without holding the release. A drawing that needs two disciplines and a QA check gets three required reviewers. A minor markup revision gets one. Either way the agent can open the session, stage the files, and assign the people, and then it stops.

When a client, an auditor, or your own QA lead pulls up this revision in three years, the approval on it carries a person's name, a date, and the exact version they were looking at.

Let the agent do the tedious work of editing all fourteen datasheets, and let an engineer review a clear diff and put their name on it.

Setup is easy: a single URL and an in-browser sign-in. It works with ChatGPT Work, Claude Code, Claude Cowork, Cursor, and any other MCP client - the steps for each are at withkord.com/mcp.