Your own AI, fully local — chat from your browser or your phone, powered by models you control.
PortableAI is a self-hosted chat interface for Ollama. New chats start as a plain Assistant — like ChatGPT or Claude — and optional named personas (a terse mentor, a patient explainer, a translator) are one click away. Chat from a dark-mode web UI or a companion iPhone app, and keep a full searchable history — all running on your own machine, with nothing sent to the cloud. Pull, browse, and update models from one place, and pair your phone over your local network in seconds.
Other 'local AI' products are chasing this from the expensive end -- Perplexity's Portable Computer (launched with NVIDIA, August 2026) needs a $4,700 DGX Spark or a high-end GPU workstation, plus an active Pro/Max subscription, and by its own admission isn't fully offline -- it calls out to the cloud for harder tasks with permission. PortableAI runs on the laptop you already have, costs nothing, and needs zero internet connection, ever -- that's not a limitation, it's the whole point.
The project lives at portableai.app. The iPhone companion is a
separate repo (portableai-ios); this tree is the server and web UI.
Packaged install — downloads the Linux AppImage or macOS disk image
into the current directory and stops. It does not launch the server.
Read install.sh first; it is the only curl-pipe installer.
curl -fsSL https://raw.githubusercontent.com/generalistcodes/portableai/HEAD/install.sh | shThat writes PortableAI.AppImage (Linux) or PortableAI.dmg (macOS) in
the current directory. Launch that file, then open the URL below.
From a clone of this repo:
python run.pyThen open http://localhost:5050. You get a chat box. Type a normal question; the reply fills in as the model writes it. Download a model from Settings if this machine does not have one yet. A phone on the same Wi-Fi pairs from Settings with the QR code or the PIN.
Prerequisites, the Linux binary, and what does not copy between machines
are in Development. Do not start ui/server.py
directly — python run.py is the supported source command.
- Streaming replies. Tokens appear as they are generated, with a blinking caret at the end of the bubble. A failure before the first token is a normal error. A drop mid-reply discards the partial text and does not save it. Details are in Features.
- Personas. New chats start as a plain Assistant. Named voices — a terse mentor, a patient explainer, a translator — are one click away in the sidebar, and you can add your own. See the Tutorial.
- Phone pairing. The server listens on the local network. A phone joins with a QR code or a single-use PIN from Settings. The plan is in Phone pairing.
- Multi-device isolation. Each paired phone sees only its own chats. The desktop, which issued the PINs, can see every device. History is searchable, and a chat can be archived, renamed, or exported.
- Cross-platform. Linux ships a complete AppImage. macOS ships an
unsigned disk image. Windows ships an executable that is not a
finished product yet. The same tree also runs from source with
python run.py. Status is in the Roadmap.
Everything past "try it" lives in docs/. Nothing below is required to
start the server.
- Tutorial — the two ways to build a persona
(Ollama CLI, and
python demo.py), example prompts, and how to add another.Modelfile. - Features — the chat UI, streaming replies, saved history, model downloads, optional update checking, and phone pairing with per-device isolation.
- Development — prerequisites,
pytestversuspytest -m integration, what copies to another machine, packaged downloads, and building the Linux executable. - API contract — every
/api/*route as implemented. Clients, including the iOS app, should follow this instead of guessing. - Roadmap — what is done and what is not.
- Phone pairing plan — QR, PIN, and why
the link is a normal
http://URL. - Model benchmark and latest results — the prompt set and the last real comparison.
Commit messages follow Conventional Commits. See CONTRIBUTING.md.
Releases are make release VERSION=vX.Y.Z.