“Your data never leaves your computer” is the strongest privacy claim a companion app can make, and also the easiest to fake. Plenty of apps say something like it while quietly syncing chats, running analytics, or routing your messages to a model on a remote server. If you want an AI companion where your data never leaves your computer, you need to know what that phrase has to mean technically, how to verify it yourself, and why only one architecture can actually deliver it. This is that guide.
What the claim actually requires
For your data to truly never leave your machine, three things must all be true at once:
- The model runs locally. If the AI that generates replies lives on a remote server, your messages have to travel there, full stop. On-device inference is non-negotiable.
- Memory is stored locally. The companion’s record of you must be a file on your disk, not a row in a cloud database.
- There’s no background egress. No analytics, no telemetry, no “sync,” no ad attribution quietly sending data out the side door.
Miss any one and the claim is false. A surprising number of “local” apps satisfy the first and fail the third, the model runs on your machine, but the app still phones home with usage data. We cover this exact failure mode in is Ollama really private.
Why cloud companions can’t make this claim
This isn’t a knock on any specific company, it’s architecture. A cloud companion’s model runs on the provider’s servers. That means your message must be transmitted off your device to be answered. The provider can encrypt it in transit and promise short retention, but the data definitionally leaves your computer. “We don’t keep it long” is a different, weaker claim than “it never leaves.” Only local inference makes the strong claim possible. For the full hierarchy of privacy guarantees, see AI and your data.
| Guarantee | Cloud companion | Local companion |
|---|---|---|
| Encrypted in transit | Usually | N/A, never transmitted |
| Short retention | Maybe (their promise) | N/A, nothing sent |
| Data never leaves device | Impossible | Yes (verifiable) |
How to verify zero egress yourself
The beauty of the local claim is that, unlike a privacy policy, you can test it.
- The airplane-mode test. Disconnect from the network entirely, Wi-Fi off, ethernet unplugged. If the companion still works perfectly, the model is genuinely local. If it stalls or asks you to reconnect, it isn’t. This is the core of the offline companion field guide.
- The packet-capture test. For the thorough: run a network monitor (like Wireshark) while you chat, and watch whether the app sends anything out. A truly local companion talks only to
127.0.0.1, the loopback address, and nothing else.
If an app passes both, the “never leaves your computer” claim is real, not marketing.
What counts as “your data” here?
Worth being precise, because “your data never leaves” should cover more than message text. A genuinely local companion keeps all of these on your machine:
- Conversations, every message, in both directions.
- Memory, the accumulated facts and summaries about you.
- Character files, the companion’s personality definition.
- Usage patterns, when and how often you chat (cloud apps often log this even when they don’t read content).
The reason local can guarantee all four is the same reason it guarantees any of them: there’s no server receiving anything. Cloud apps frequently protect message content while still collecting metadata and telemetry, which is its own leak, since when and how often you talk to a companion is sensitive too. Local sidesteps the whole category by never transmitting in the first place.
How to get a genuinely local companion
The DIY foundation:
-
Install Ollama:
curl -fsSL https://ollama.com/install.sh | sh -
Pull a model sized to your hardware and chat at
127.0.0.1:11434. -
Keep memory and characters on local files, and confirm zero egress with the tests above.
The one nuance to watch: some tools offer remote/cloud model options that quietly break the guarantee, stick to models that run fully on your hardware. (Ollama’s own -cloud variants are the classic trap; we flag it in is Ollama really private.)
One honest note on shortcuts: Ember meets this article’s bar rather than dodging it, she runs on your own GPU and sends nothing to a server. The difference is that she arrives assembled, for $29 paid once. That’s a real posture, but it isn’t zero egress. If the claim you need is “my data never leaves my computer,” build the local stack above, it’s the only architecture that can make that claim, and you can verify it yourself.