The AI that lives on your Mac, knows your entire firm, and keeps your data 100% private*
* Made you look. No, really. 100% private. Nothing ever leaves your Mac.
Ask your practice management software a question. A real one. Not “run the aging report.” A question. “Which clients got slower at paying me this year?” “What kind of work actually makes me money?” “What am I forgetting?”
Silence, right? Twenty-plus years of legal software and the deal never changed: you do the data entry, and if you want answers, you build the report yourself, export the spreadsheet, and squint.
Oasis is a new deal. It’s an AI built into TimeNet Law that has read your entire firm. Every client. Every matter, time entry, invoice, expense, and trust transaction, going back as far as your data goes. Ask it a question in plain English and it answers with your actual numbers. It drafts, it classifies, it reconciles, it designs flat-fee packages from your own billing history, and it will write you a briefing on the state of your firm while you sleep.
And it does all of this on your Mac. Not “your account.” Not “our secure cloud.” The aluminum on your desk. Turn off your Wi-Fi and ask it anything. It won’t even notice.
That last part is the entire story. So let’s tell it.
The Weekend Version
I’ll let you in on an industry secret: bolting a chat window onto an app takes about a weekend. Sign up for a cloud AI service, get an API key, wire up a text field. Done. It demos great, investors love it, and over the next year you’re going to watch nearly every legal software company on earth ship exactly that, each with a cute name and a press release leaning hard on the word “secure.”
The weekend version has a problem, and it isn’t small. Every question your firm asks travels to a server you’ve never seen, owned by a company you have no relationship with, governed by a privacy policy that changed while you were reading it. Client names. Matter details. Dollar amounts. Settlement postures. Confidential documents. The exact material you’re professionally obligated to protect.
Using a cloud legal billing system or cloud legal AI is like buying a safe for your files and taping the combination to the door. The company that sold you the safe promises they’ll never peek. Their privacy policy reserves the right to. Rule 1.6 did not get an AI exception. Your vendor’s “enterprise agreement” is not an ethics opinion. And “Zero Data Retention” policies don’t guarantee you anything.
I’ve written before about the bet I made 23 years ago: your data lives on your machine, in files you own. For most of those years that architecture was unfashionable. Then AI arrived and asked the entire software industry one question. Where does the data live? And suddenly the unfashionable answer was the only good one.
So I didn’t build the weekend version. I spent years building the other one. An AI that runs entirely on your Mac, where the privilege already lives. Nothing to intercept. Nothing to subpoena from a third party. Nothing to leak.
Not a promise. An architecture.
This wasn’t easy. But easy and worth it rarely travel together.
Your Firm Doesn’t Fit Through the Door
A language model reads text as tokens (word-chunks, roughly), and even a big model can only hold so many in its head at once. Its “context window” is finite. A law firm’s history is not. Twenty years of matters, entries, and invoices runs to millions upon millions of tokens. The model’s window is a mail slot. Your firm is a filing room.
The weekend version solves this by not solving it. Ask a bolted-on chat window about “my top clients” and it sees whatever scraps got stuffed through the slot, which is why those bots so often answer with the confidence of a first-year associate who read one folder. (We’ll get to the confidence problem. Oh, we’ll get to it.)
Oasis solves it with an engine that slices your firm into digestible pieces and knows which pieces matter for the question at hand. Ask about one client and it assembles that client’s full picture. Ask a firm-wide question and TimeNet Law computes across everything first, then hands the model a summary it can actually hold. And for the heavy jobs, Oasis runs a multi-phase pipeline: read the firm in batches, find the patterns, consolidate them, design the recommendations, then write the report. Discovery. Consolidation. Design. Report. Dozens of passes, one coherent answer. And it’s fast.
Getting those batches right was one of the longest fights of the whole project. Pack too little into each pass and a 300-matter analysis takes all night. Pack too much and the model drops key details and confidently hallucinates the missing pieces. I tuned the batching until it used the window like a moving truck instead of a mail slot and cut the number of passes by 4x.
(macOS picked a fight of its own here. The system kills an idle network connection at 60 seconds, and a deep analysis “thinks” in silence for longer than that before the first word arrives. The fix involved throwing out the polite networking layer entirely and building my own. Details matter.)
One Model, One Floor
Early builds of Oasis had model tiers. Three different AI models, picked automatically by how much memory your Mac had. More RAM, bigger brain. It was clever, and I hated it. It meant two attorneys could ask the identical question and get different-quality answers, and answer quality is not a place I’m willing to be clever.
So I drew a line. One model for everyone, and I’d find the smallest Mac that could hold it. The answer, after months of testing: a 26-billion-parameter model and a 24GB memory floor. Every feature in Oasis was built and tested against that floor. If it didn’t run beautifully on a 24GB MacBook Air, it didn’t ship.
Why not smaller? I tried. And tried. Sometimes it seemed promising, but fighting an LLM that’s too small to do what you need is like swimming upstream. Clever guardrails. Aggressive tokenizing. Endless prompt iterations. Only to be caught in the same hallucination trap.
Here’s a dirty little secret about AI. It’s designed to answer your question. It wants to please you. And when it doesn’t have the context available to provide the correct answer, it straight up fabricates one. And suddenly you’re seeing more hallucinations than Woodstock.
Legal reasoning on smaller models just wasn’t going to work.
Okay, but what if you’ve got 64GB of RAM? Why would you possibly want a model that only takes 24? Why not bigger? Well, I tried that too. A 109-billion-parameter monster lived in the dev cycle for a while. And put simply, it didn’t earn what it cost. It’s slower. Way, way slower. What Oasis answers in about ten seconds took minutes. It runs hotter. It destroys battery life. And the answers it finally spit out weren’t better. Sometimes they were worse.
The model I shipped is the one that earns its seat. It turns out a moderate model with finely tuned guardrails, prompt engineering, anti-psychedelic defenses, and your firm’s actual records in front of it runs circles around a frontier genius locked out of the filing room. And that’s the whole problem with the bolted-on cloud chat window: it gets all of your secrets and none of your context.
Getting to that floor meant sweating things nobody puts in a press release. My favorite: the memory an AI needs isn’t just the model, it’s the workspace the model uses to read long documents, and during development I watched one lazy default try to grab 42 gigabytes of it to read a single contract. The Mac sat there for minutes, warming the room, thinking about thinking. Oasis now sizes that workspace dynamically, document by document. You will never see it happen. That’s the point.
The Routing Brain
Language models are dreamy readers, but someone has to decide what lands on their desk.
When you ask Oasis a question, the model isn’t the first thing that touches it. First, a classifier I built in plain, boring, testable code figures out what kind of question it is. A client question? A matter question? A firm-wide number crunch? A compliance scan? A drafting request? Then it gathers exactly the right context, runs the actual math, and only then does the AI get involved. With the real data as its anchor.
Think of a great paralegal. The quality of the partner’s answer is mostly decided before the partner starts thinking, by what got pulled and stacked on the desk. That’s the classifier’s job. I iterated on it relentlessly, because when the routing is wrong, the AI answers the wrong question beautifully. Like a Pulitzer Prize-winning author crafting a brilliant biography about the wrong person.
The War on Psychedelics
Now the uncomfortable part, the one the industry would love to skip. Every AI hallucinates. Every single one, at every price point, no exceptions. Ask a model for a number it doesn’t have and sometimes it will just invent one, delivered in a confident, fluent, entirely wrong sentence. In most industries that’s embarrassing. In yours it’s malpractice-adjacent.
You cannot fix this by adding “please don’t make things up” to the prompt. (I tried. Everyone tries.) You fix it with architecture. Oasis is built on one operating assumption: the model lies, and the system’s job is to make sure it can’t get away with it.
The model never does math. Every dollar figure, every hour count, every percentage in an Oasis answer is computed by TimeNet Law, in code, from your actual records, before the model ever sees it. The AI narrates the numbers. It does not calculate them. It’s the storyteller, not the accountant.
Anything touching your data runs deterministic. Creativity is literally a dial on these models. For anything data-driven, I turn it nearly to zero and pin the randomness to a fixed seed (it’s 42, naturally). Same question, same data, same answer, every time. Boring. But boring and correct beats colorful and wrong every time.
Structured data is grammar-constrained. When Oasis needs the model to produce data instead of prose, the model is physically constrained to a schema. It cannot produce a malformed answer. I didn’t ask nicely. I removed the option.
Nothing changes your records without your sign-off. When Oasis proposes actions, they land in a review window and you approve them line by line. The AI proposes. You dispose. And every change is flagged “via Oasis” so it’s transparent, and infinitely undoable in your Event Vault. Even something done a week ago can be independently rolled back. Not a full database restore. Nothing lost. Just a quick escape hatch if you need it.
Are hallucinations extinct? No, and anyone who tells you theirs are is selling the weekend version and lying to your face. But I spent months hammering them down. Measure, tighten, constrain, re-test. Wash, rinse, repeat, hundreds of times, until the psychedelics wore off and the answers stood up. The result is an AI I trust in front of attorneys. That’s the highest bar I’ve got.
So What Does It Do?
I’ve been telling you how the watch was made. Now let’s talk about telling time.
Oasis is a crew, not a chatbot. Six personas, each with the same deep knowledge of your firm and a different job: First Mate (your everyday analyst), Lookout (risk), Rainmaker (revenue), Harbormaster (trust and compliance), Navigator (operations), and Wayfinder (strategy, the one that designs flat-fee packages). Ask from the chat window, or by voice, or right-click nearly anything in TimeNet Law and hand it to the crew.
Oasis is $199, one time. Not per seat. Not per month. Not per token. (The AI industry is about to teach your profession the word “metered.” When the bubble pops, you’ll be glad I went the other way.) No signup, no account, no usage dashboard. You own it, like you own everything else in TimeNet Law.
And rather than tour every feature, let’s do something more fun. If you’ve got TimeNet Law and a Mac with 24GB of memory, here are seven ways to spend your first night.
Seven Things to Try Tonight
1. Ask how the firm is doing. Type it just like that. “How’s the firm doing?” You’ll get revenue, collections, work in progress, and the trendlines underneath them, from your actual data, in plain English. Are things getting better or worse? Now you know at a glance, instead of after an afternoon of report-building. It’s a handshake. Shake it.
2. Ask what you actually earn per hour. You know your billing rate. Do you know your collected rate? After the write-downs, the write-offs, and the slow payers, the number on your engagement letter and the number that reaches your account are two different figures, and most attorneys have never once seen the second one. Ask Oasis for your effective rate. Sit with it. That number changes firms.
3. Ask Lookout what’s slipping. Stale matters nobody has touched. Unbilled work aging quietly in a drawer. This is the question you’d ask at 2 a.m. if software could answer it, and now it can, at a civilized hour. Unbilled time is money you already earned and simply haven’t asked for. Ask Oasis to go find it.
4. Let Oasis bill and catalog an expense. Snap a photo of a receipt and drop it into the matter window. Watch it become an expense entry, vendor, amount and date filled in, ready for your eyes. You approve, and the receipt gets billed, image renamed and filed in your Receipts folder. Five seconds. It’s a party trick, except the party is your month-end billing and the trick is real. And when the audit comes due, that record has your back.
5. Let it organize your practice areas. If your matters have never been tagged by practice area (no shame, nobody’s have), ask Oasis to classify them. It reads every matter and proposes an assignment for each one in a review window where you approve or fix each line. Hundreds of matters, organized in minutes, and nothing written without your say-so. Then ask question #1 again and watch the answers sharpen by practice area.
6. Ask Wayfinder to design a flat-fee package. Open a matter you’ve handled a dozen times and ask Wayfinder for a Blueprint. It reads what that work has actually cost you across your own history (the hours, the overruns, the outcomes) and designs a flat-fee package: scope, pricing, timeline, ready for you to edit. And when you’re ready to go bigger, Strategic Analysis reads the whole firm and designs a catalog. This is the feature I’d put up against anything in legal tech, because it’s impossible without both halves in the same room: an AI tuned for truth, and decades of your billing history. Nobody else has both. Nobody else can.
7. Go to bed. Really. Schedule the Rainmaker Briefing, close the laptop, good night. While you sleep, Oasis reads the state of your entire firm and writes you a briefing for the morning: what moved, what needs attention, where the money is. No server did this for you. Your Mac did, at your desk, with the office door locked. Everyone has scheduled reminders. Only TimeNet Law has scheduled intelligence.
Where the Water Comes From
A few weeks ago I introduced this era of TimeNet Law with a line I’d been carrying around for years: a glass of ice water in hell.
Let me be precise about the metaphor now, because I’ve thought about it since. The hell is real. Rented software, mined data, metered everything, support tickets sent into the void. And an oasis isn’t a mirage, and it isn’t a miracle. It’s just the place where the water turns out to be real, and nobody is charging you by the sip.
The gold rush is coming for your profession. Over the next year you’ll be offered a thousand chat windows, and every one of them will demo beautifully. Ask each one a single question before you type a client’s name into it: where does my data go?
Oasis has a one-word answer. Nowhere.