Local AI.
Transcription.
Private Mesh.

Yes. MyOwnLLM is open source under the MIT licence, which means you can use it forever, even commercially, even offline, even after this project is no longer maintained. There's no premium tier, no account, no telemetry by default.

No. On first run, MyOwnLLM measures your machine and downloads a model that fits, from edge-class on a small Pi up to a big model on a workstation with a 5090. When a better model comes out for your hardware, it swaps it in. When the app itself needs an update, it stages it on launch and applies it next time you open the app. You don't have to choose, configure, or babysit anything.

No. MyOwnLLM runs fine on a single computer. It picks a model your machine can handle and you're off. The multi-device part kicks in when you add a second one: each device joins the network, the AI can route inference to whichever has the right power, and the conversation follows you between them. Start with one. Add the rest when you feel like it.

Yes. If you've got MyOwnLLM on a machine at home and another on the laptop you're travelling with, they find each other through your private mesh and you can talk to the home model from the road. Your laptop's chat, your workstation's GPU. Transcripts and conversations recorded on one device are reachable from the others. The connection between them is peer-to-peer and encrypted; there's no MyOwnLLM-operated server in the middle. Run a local relay on your LAN if you'd rather not touch the public internet at all.

Yes, after the first download. The first time you run it, MyOwnLLM downloads the AI model itself (typically 3–15 GB). After that it runs entirely offline. You only need internet again if you want to update the app, switch to a different model, or find peers on a public Nostr relay (a local-network relay works too if you'd rather not touch the internet for that).

No. MyOwnLLM works on a stock Mac, a regular Windows laptop, even a Raspberry Pi 4 with 2 GB of RAM. It picks a model your machine can actually run. If you do have a GPU it'll happily use it to go faster. And if you have a beefy GPU in another room, your smaller devices can borrow it over the mesh.

Only when you're actively chatting. And even then it's bounded by what your computer can comfortably handle, because the model is chosen to fit. Idle, MyOwnLLM uses no meaningful power.

The app itself is small (under 50 MB). The model it downloads on first run is the big bit: typically 3 GB on a Pi or small laptop, up to about 15 GB on a powerful machine. Models you stop using are cleaned up automatically.

Everything runs on your devices. Chats, transcripts, files: they never leave the machines you own. When devices on your network talk to each other, the connection is peer-to-peer (WebRTC) and mutually authenticated; the only thing a public Nostr relay sees is the encrypted handshake.

The chat doesn't just answer. It can do things. It has tools to manage your devices (add a Pi, switch networks, approve a peer), read files, write files, and run shell commands. Anything that touches your shell or writes to disk pops up a permission prompt: Deny, Allow once, Always for this specific command, or Always for the tool. Your choices are saved per device, so each machine has its own policy.

For everyday things, yes. Drafting emails, summarising articles, answering questions, looking at photos, helping with writing or code. It's not as smart as the very largest cloud models, but it's a lot smarter than nothing, and it's yours. As local models get better, MyOwnLLM updates itself to use them.

That's the point. Install it everywhere. Each machine gets a model that fits it, and they join the same AI. The Pi runs a small nimble model; the gaming PC runs a big clever one; questions you ask on any device can be answered by whichever has the right power for the job.