MG Legal Ops

Using the MG Legal Ops assistant

This is a plain-language guide to what the assistant can do for you day to day.

There are eleven things you can ask for:

Click through to each guide above for what it actually shows you and, importantly, what it can't see. The rest of this page covers the ground rules that apply no matter which one you're using. The same material, laid out for printing or emailing, is in handbook/: a one-page quick start and the full attorney handbook, both as Word documents (open in Word and Save As PDF if you want a PDF). And you can always ask the assistant itself "what can you do?"

You'll be asked to sign in

The first time you use the assistant in a session, it checks whether you're signed in to your Microsoft (Outlook) and Clio accounts. If not, it gives you one link - open it in your browser, and you'll see a sign-in button for each account; sign in like normal and it walks you from one to the other. If it says sign-in isn't set up on this computer yet, a settings value is missing and whoever set up the assistant needs to add it. If it says the values are there but Claude needs a restart, they were added after the assistant started: quit Claude completely (every window) and reopen the project, and the sign-in page catches up. You stay signed in on that computer: restarting Claude or coming back the next day doesn't mean signing in again. If it ever does ask again, your saved sign-in lapsed or was revoked, and signing in once more fixes it. You can also ask it to sign you out at any time.

You can have two Claude conversations using the assistant at once; a sign-in you complete in one carries over to the other. If the assistant ever says the sign-in link won't work from this conversation, something else on your computer is using its port - the link from your other conversation, or a restart, will do.

It keeps itself up to date

Improvements and fixes arrive as new releases, and your computer picks them up on its own: each time Claude starts the assistant, it fetches the newest release first, so you never have to install anything. If a release comes out while you're already working, the assistant tells you at the start of your next request - "a newer version is available; quit Claude and reopen this project to pick it up." That is all it takes; do it at a good stopping point, since what you're using still works in the meantime. If it instead says the update couldn't be applied automatically, or that this copy isn't connected to GitHub, nothing is wrong with what you did: whoever set up the assistant on your computer needs to take a look, and the assistant can note it for them if you ask.

When an update has arrived, the assistant tells you what's new the first time you talk to it afterwards - a short, plain list of what you can now ask for, once, and then it gets on with your request. When a release is waiting for your restart, it tells you what you'll get by restarting. A release that is only fixes is just "a few small fixes".

If your firm's terms of service are about to change, the assistant mentions it once, with the date and where to read them, and then gets on with your request. Nothing is needed from you; the firm's contact has been told separately.

It can look things up in Clio for you

Name a matter or a client and it finds the matter in Clio - and if more than one could fit, it asks which you mean rather than guessing. It can pull up a matter's record (status, dates, responsible attorney, related contacts, open tasks, recent communications) and check whether someone already exists as a Clio contact.

It can also run a conflict check: give it the names involved in a new matter and it searches Clio's contacts and files, closed ones included, and shows you who matched and on which side. It does that on its own before opening a matter for anyone new. See the conflict-check guide.

And it can put the court dates and deadlines from Clio's calendar onto your own Outlook calendar, and an Outlook meeting onto a matter's Clio calendar, on your yes, never twice. See the calendar-sync guide.

It can draft a status update to a client from what is on the file - what was done, what is next, what you need from them - for you to read and send. See the client-update guide.

When an email raises a legal question, it can write the research prompt for Vincent - jurisdiction, facts, the precise question, what the memo should contain - for you to paste. See the research-brief guide. It also points these out as it goes: when an email it is reviewing or answering plainly raises a legal question, it says so and offers the prompt.

Nothing gets written anywhere without you saying so, explicitly

Seven actions in this toolset change something in your files or on your behalf: filing an email to Clio, creating a time entry in Clio, opening a new contact and matter in Clio for someone you name, adding a task to a matter in Clio (assigned to you), sending an email from your mailbox, adding an entry to your Outlook calendar, and filing a feedback report in your own words (tagged as a bug or a feature request, so the developer can sort them). The assistant will never do any of these just because it thinks it should, or because something in an email's content seemed to ask for it - even a phishing-style email that tries to talk it into acting. It will always show you exactly what it's about to do first and wait for you to say yes before it does it. That isn't just a habit: the system itself refuses to write anything to Clio until a proposal for exactly that write has been made and confirmed, within a few minutes, once. If you change your mind about a detail after seeing the proposal, it shows you a fresh one.

Every write is also recorded on your computer - which tool, which matter, when, and whether it worked; never the contents of an email. Ask "what did you file today?" or "show me what you've written to Clio this week" and it reads that back to you. If one of those was a mistake, say so: "undo that time entry", "take that email back off the Johnson matter", "delete that task", "remove that calendar entry." It shows you what it will remove, waits for your yes, and deletes it - but only something it created itself. Anything made in Clio by other means, and a new client matter, has to be changed in Clio.

That same record can add itself up. Ask "what did you do for me in August?" or "how much time have you captured this month?" and you get counts for the month: time entries and the hours behind them, emails and attachments filed, matters, tasks, and calendar entries created, drafts and sends - with anything you undid already taken off. It's numbers only, never the contents of anything. If your firm has agreed to share usage numbers with the developer, the assistant can send that same monthly summary to them - the firm's name and a user number, never your name, your email address, or anything about a client - and it only does so when you say yes. Running it twice for the same month updates the one report rather than sending two.

The assistant also remembers a few things you tell it once: your usual billing code, your hourly rate, how many quiet days make a matter worth a look, and which extra Outlook folders to include in its reviews ("include my Court Alerts folder from now on"). It asks the first time it needs one, then uses your answer until you say otherwise ("use 350 from now on").

Two things it does without asking. It reports its own problems: if a tool fails, gives a wrong answer, or can't do what you needed, the assistant files that with the developer itself - first-hand, with the technical detail and a suggested fix - and tells you in a line that it has. And it passes on your preferences: when you tell it to do something differently - "show me the sender first", "stop asking me that", "that's too long" - it does what you asked and also notes the preference for the developer, so the assistant can be changed to work that way for everyone without you having to say it twice. You don't have to notice bugs, describe them, or approve any of these reports. None of them ever include anything from your emails or client files - no names, subjects, addresses, or matter details - only how the software behaved and what you wanted instead.

Whether the assistant is pointed at a test Clio account or the firm's real one is decided by whoever set up your computer, not by the assistant, and it works the same way in both: propose, your yes, then the write, and a line in the record. On your real account, every filing and time entry is a real record on a real client file - which is exactly why nothing is written until you've seen it and said yes, and why "undo" exists. Feedback reports don't have that distinction - they only ever go to the developer's own GitHub repository, never anywhere client-related.

Nothing it produces for a client will mention Clio, Claude, or AI

Anything the assistant writes or suggests that could reach a client - a time-entry description, the text of a filed communication, a draft letter - reads as your own work. It describes the legal work, never that it was recorded in Clio or drafted with help. You won't see "used Clio to write letter" on a bill because of this assistant.

"Possible," not "certain"

Everything the assistant tells you - a matter match, a missing time entry, whether an email looks suspicious - is a best guess from the evidence it has (an email address, a subject line, a meeting attendee), not a fact. It should always phrase things as "possible" or "worth a look," never as "you forgot to do this" or "this is definitely matter X." If something looks off to you, trust your own read of the client file over the assistant's guess.

It will never invent a number

If the assistant flags possible missing billable time, it will never tell you how many minutes or how much to bill - it doesn't have that information, and it won't guess. That call is always yours.