How to Let ChatGPT or Claude Do Legal Work Without Losing the Audit Trail

Your firm almost certainly pays for an AI assistant already. Lawyers use ChatGPT or Claude for research, summaries, and first drafts, usually by copying text in and copying answers out.

That copy-and-paste habit is the problem. It works, but it leaves no record of what the AI did, and it sends client text into a personal chat account. There is now a better way: connect the assistant your firm already uses directly to a legal work platform, so the AI can do real work on your matters while every action it takes is recorded.

This guide explains how that connection works, in plain language, and how to introduce it at a firm step by step. No technical background needed.

First, the problem with copy and paste

Today, most AI-assisted legal work happens in a loop like this:

1. A lawyer copies contract text into a chat window.
2. The AI suggests changes.
3. The lawyer pastes the results back into Word and tidies up by hand.

Three things go wrong.

The record is lost. Nothing in the final document shows which language the AI proposed and which language the lawyer wrote.

The AI is working blind. It sees the paragraph you pasted, not the rest of the agreement, the earlier drafts, or the other documents on the matter.

Client text went somewhere. A personal chat account keeps history under terms nobody at the firm has reviewed. That is a confidentiality question, not a technicality.

What is MCP? The plug that connects your AI to other tools

MCP stands for Model Context Protocol. Behind the acronym is a simple idea: it is a standard plug that lets an AI assistant connect to another piece of software and use it, with your permission.

You have already used this pattern elsewhere. When you let your calendar see your email, or let a document signing tool access a file, you clicked a button that said something like "Allow access." MCP brings the same idea to AI assistants. ChatGPT and Claude both support it, and both call these connections "connectors."

Once gitmatter is added as a connector, your assistant can see your matters and use gitmatter's tools: open documents, review them, suggest redlines, pull key terms into a table. You ask for things in plain English. gitmatter does the actual work and writes every action into the matter's history.

Two roles, clearly split:

• The assistant (ChatGPT or Claude) is the colleague you talk to. It understands your request and decides what to ask gitmatter to do.
• gitmatter is the file room and the record. Documents, matters, playbooks, and clause libraries live there. Every change is saved with who made it, what changed, and why.

Because the record lives in gitmatter and not in the chat, it does not matter which assistant your firm uses this year. Switch from ChatGPT to Claude, or use both, and the matter history stays complete.

How to connect: about five minutes, no code

You connect once, and it works from then on. The steps are the same kind you follow to link any two apps.

If you use ChatGPT

1. In ChatGPT (paid plans), open Settings, then Connectors.
2. Add a custom connector and enter your firm's gitmatter address: https://your-gitmatter-address/api/mcp.
3. ChatGPT opens a gitmatter sign-in page. Sign in and click approve.

That is the whole process. There is no code to copy.

If you use Claude

1. In Claude (desktop or web), add a connector.
2. Enter the same address: https://your-gitmatter-address/api/mcp.
3. Sign in to gitmatter and approve.

The sign-in step matters: the assistant acts as you, with your access and nobody else's, and everything it does is recorded under the assistant's name inside your account. Technical colleagues who use command-line tools can connect too; the connect-an-agent guide covers those.

What you can ask for once connected

Once connected, you brief your assistant the way you would brief a junior colleague:

• Questions about a matter. "Summarize the open issues on the Meridian acquisition." The answer comes from the matter's own documents, with citations to the exact passages, so checking it takes one click.
• A table from a pile of contracts. "Ask these eight questions of every contract in the data room and give me a table." Each cell links back to the passage it came from.
• A first-pass markup. "Redline this MSA against our standard positions." The suggestions arrive as tracked changes with the reasoning attached, each one waiting for a lawyer to accept or reject.
• A draft from your templates. New documents start from the firm's templates and approved clauses, not from whatever the AI imagines a contract looks like.

The pattern in every case: the AI proposes, a named lawyer decides, and the record shows both.

The safety rules, built in rather than promised

Letting an AI touch matters is only responsible when the safeguards are part of how the system works, not lines in a policy document.

1. Nothing skips the record. Every change the AI makes goes through the same history as a human edit. There is no side door.
2. The AI has your access, not master access. It acts as the person who connected it and can only see the matters that person can see.
3. AI edits are labeled as AI edits. The history shows the change came from the assistant, under which lawyer's account, with what reasoning.
4. Client text stays under the firm's terms. gitmatter's own AI features run on the firm's own account with the AI provider, set up so documents are not stored or used to train anyone's model.
5. Everything can be undone. Any change the AI made can be found in the history and reversed on its own.

A four-week way to introduce it

Do not start with your hardest matter. A rollout that works:

1. Week one: questions only. Connect the assistant and ask it about a closed matter. The team learns what the AI can see and how citations work. Nothing changes in any document.
2. Week two: extraction. Run a table across a stack of documents. Lawyers check the answers cell by cell. Still nothing changes in the documents.
3. Week three: supervised redlining. Let the AI make a first-pass markup of a live agreement against your playbook. A lawyer accepts or rejects every suggestion.
4. Week four: read the record together. Open the matter history as a team and read what the AI did. This is the step that converts skeptics: the "black box" turns out to be a readable list of changes.

The record outlasts the assistant

AI assistants change fast: new models, new vendors, new features every quarter. Your matter records should not change with them. That is the real argument for this setup. The firm's work, its history, and its accountability live in a system the firm controls, and whichever assistant you prefer this year simply plugs into it.

gitmatter is that system: an audited legal backend that ChatGPT and Claude connect to, with a complete history under every matter. Book a demo with your own assistant and watch the record build itself.