Privacy architecture: what leaves your Mac, and what never does.

Every AI product page includes the sentence "we take your privacy seriously." It costs nothing to write and nothing to break, which is why everyone writes it. I want to do something more useful and publish the inventory instead: every category of data OMEGA handles, whether it leaves your Mac, where it goes when it does, and whose name is on the account it travels under. If a claim in this post is wrong, it is checkably wrong. That is the point of writing it this way.

The short version: your work lives on your Mac. Model calls go to the providers you configured, with your keys, and to nobody else. Our servers see licensing and billing traffic, never content. And if that is still too much network for your situation, there is a local-only posture where a model call never leaves the machine at all.

What never leaves

OMEGA is a native macOS app on Apple Silicon, and the workspace is your machine, not a remote sandbox rendered into a window. Everything the product accumulates about you is stored locally:

• Conversations. Every chat, on every surface, sits in a local database on your Mac. There is no server-side transcript of anything you have ever typed into OMEGA.
• Files and projects. Project Canvas can import a whole local folder or a git repo, and the import stays on your disk. The same goes for Design Studio: your canvases, frames, and design systems are local files, not documents in our cloud.
• Memory and the Brain. The 4-layer memory and the Brain knowledge graph, the parts of OMEGA that compound in value the longest, live on your Mac in full. What OMEGA knows about your projects and preferences is not mirrored to us in any form. We wrote about why that accumulation matters in our first post.
• Operational state. Budgets, per-workspace API keys, agent configurations, and the per-app approvals that govern computer use: all local.

Two things people assume phone home and do not. Voice runs on-device, so raw audio is processed on your Mac rather than streamed to a hosted transcription endpoint. And the orchestration state of OMEGA's Neural-Fractal Agentic AI™ (NFA) runtime, the task trees and routing decisions it builds while it works, is local state too, not a server-side artifact of how you think.

What leaves, case one: model calls to providers you chose

Here is the part most local-first marketing skips. Unless you run fully local, the context for a turn goes to a model provider, because that is what an API call is. If you route a heavy turn to Claude Fable 5, the prompt OMEGA assembles for that turn, including whatever file excerpts and memory slices the runtime judged relevant, is sent to Anthropic. There is no version of "we use frontier models" that avoids this. Not from us, not from anyone.

What architecture can control is who the call goes to, whose account it travels under, and what else rides along. In OMEGA:

• Model calls go only to providers you explicitly configured. There is no default provider and no silent fallback to a vendor we picked for you.
• Calls go direct from your Mac to that provider. We run no relay or inspection point in the middle. This is also why BYOK carries no markup on tokens: we are not in the path, so there is nothing for us to meter.
• Your API keys are stored on your Mac and used from your Mac. Subscription routing to Claude and Codex works the same way: your subscription, under your own account and agreement with that provider.
• Keys are per-workspace. A client engagement can run entirely on the client's key, so their material never rides on your personal account, and your other workspaces never ride on theirs.

The practical consequence is that your provider relationship stays yours. What a provider retains and trains on is governed by the agreement between you and them, on the tier you chose, not by a bundled reseller arrangement you never read. If you negotiated zero retention with a provider, OMEGA does not undermine it, because we never see the traffic.

One more thing can leave your Mac, and only by explicit choice: an Automation you deliberately place on the always-on VPS runner so it keeps firing while your laptop is closed. That is a decision you make for a specific Automation with the tradeoff in plain view, not a default we chose for you.

What leaves, case two: licensing and billing, never content

OMEGA does operate servers, and I want to be exact about their job. They validate your license and handle billing. That traffic carries account identity and entitlement state: who you are and what you are entitled to run. It never carries content. Not conversations, not file contents, not memory, not the Brain.

This is a structural separation rather than a policy promise. The licensing plane and the content plane are different pipes to different places: entitlement checks go to our servers, and model calls go to your providers. Everything else stays on your disk. A policy can change with a terms-of-service update. A pipe that was never built cannot quietly start carrying your conversations.

When the workspace is the vendor's computer

Now the contrast, because the products on the other side are good.

ChatGPT's memory is genuinely useful, and Projects are a sane way to scope context. But every fact those features learn about you is stored in your OpenAI account, on OpenAI's servers, and works only there. The workspace and the vendor are the same object. End the relationship and the accumulation ends with it.

Manus is the purest version of the pattern: a cloud agent whose working environment is a vendor machine. The My Computer desktop app it shipped in March reaches down to your desktop, but the agent's home remains the vendor's cloud. Manus also happens to be the cleanest illustration of why custody matters. Butterfly Effect, the company behind it, was acquired by Meta last December, and every workspace its users had built changed owners overnight, with no action and no signature from any of them. I have no reason to believe Meta will mishandle that data, and that is not the argument. The argument is that data you do not hold can change hands without you.

Claude Cowork deserves the most careful treatment, because Anthropic got something real right. Cowork, still a research preview, works on actual files in folders you choose, with the same agentic core as Claude Code, and for non-technical knowledge work that chosen-folder design is meaningfully better than an upload box. Scoping an agent to explicit directories is the correct instinct, and I will say so plainly. But the reasoning loop still runs in Anthropic's cloud, under Anthropic's account structure. Everything Cowork reads in order to do its job becomes model context on the vendor's side of the wire, and the running record of your work with the agent accrues there too. The files are yours. The brain that read them is not.

And the other side of the ledger: these vendors run serious security teams, better resourced than most companies' entire IT departments. A cloud workspace follows you to any device with nothing to install, and backup is somebody else's job. If your main threat is a stolen laptop or a spilled coffee, the vendor cloud is arguably the safer place for your data. Local custody answers a different threat model: acquisitions, subpoenas, policy drift, training appetite, and the plain fact that terms you accepted in January can be different terms by December.

Local-only, for the strictest posture

For some work, "your keys, your providers" is still too much network. Client material under a strict NDA, or a legal matter where the retention questions have teeth. For that, OMEGA runs local models through Omega-MLX on Apple Silicon, and you can point every tier of the runtime at them. In that posture a model call never becomes a packet. The prompt and the answer both happen on the machine in front of you.

Now the cost. Local models are smaller and weaker than the frontier, and the gap is real on hard reasoning. A local model with your full memory and Brain behind it is still more useful than people expect, because most turns in a working day are not frontier-hard, but the tradeoff is real. You are trading capability for custody, and for some material that trade is exactly right.

OMEGA is not the only project taking local seriously. OpenClaw is local-first with an open skill registry. Hermes Agent is self-hostable and shipped a native desktop app this month. Ollama and LM Studio have made running open models on a Mac almost boring, in the best possible sense. Assemble your own stack from those parts and you will land somewhere philosophically close to us. What OMEGA adds is that local models drop into the same runtime as everything else, so they inherit the full 4-layer memory and Brain plus the 200+ agents and 5,300+ bundled skills, instead of starting from nothing.

Do not take my word for any of it

The pleasant thing about a native app is that its network behavior is observable. Put Little Snitch or LuLu in front of OMEGA and watch what it opens. You will see model traffic to the providers you configured and licensing traffic to us. That is the whole list. Switch to local models and unplug the network, and OMEGA keeps working, because nothing it needs lives anywhere else.

The fuller technical write-up is on our security page, and the plain-language version of what we collect and why is on the privacypage. Read them next to any competitor's equivalent pages and compare the level of specificity. That comparison is the argument.

Where your data sleeps is an architecture decision, and it gets made for you long before you read anyone's policy page. We made this one the way we wanted it made for ourselves: the work stays home, and the models come to it.