Portico · local AI

What happens
to what you type?

Portico runs large language models locally on your own computer — a private, offline alternative to ChatGPT. Your words go to a program on this machine and come back. One feature can reach the internet, and this page tells you which.

Download for Windows macOS and Linux
  • FreeMIT, no plan to upgrade to
  • No accountnothing to sign up for
  • 34 models1 GB to 42 GB, built in
  • No telemetrynot a single event
Portico as it opens: a sidebar with chats and projects, a message box, and a line reading Runs 100% locally.

The whole app on opening. No wizard, no account, no key to paste.

The route your words take,
and everything that leaves.

Cloud assistants add a third hop you cannot see. Portico has no server to add one to — and the exceptions are listed below, in full.

Your words go from the window to an engine on the same computer and back. Only web search crosses out, and only if you turn it on. YOUR COMPUTER You type the Portico window The engine llama.cpp, local back and forth A search engine off by default only if you switch it on
Nothing crosses the box unless you ask it to, and the app says so on screen while it does.
The same list the app shows in its own Settings.
What goesWhenHow much
Nothing Chatting with a local model
Your words Web search, off by default
Your address Downloading a model
Your address The update check, which has a switch
Everything Only if you use someone else’s engine
Portico's Settings, showing the detected machine, the memory budget, and the list of what leaves the computer.
The same list, inside the app.

On disk: chats and keys encrypted with your system keyring. Protects a stolen laptop. Not someone already logged in as you.

Getting started:
install, pick, type.

The engine ships inside the app. Nothing to wire together.

  1. Install it. One file, no administrator rights.
  2. Pick a model. Thirty-four built in, 1 GB to 42 GB. Portico measures your machine and marks which fit.
  3. Type. First reply in a second or two.
A conversation in Portico: a packing list answered by Llama 3.2 3B, with the speed and context usage along the bottom.
A real answer from Llama 3.2 3B on an ordinary laptop — 48.6 tok/s, and nothing left the machine.
Portico's model catalogue: each entry shows its size and a green Fits badge, with Installed marked on the ones already downloaded.
Portico measures the machine it is on and marks every model Fits, Tight or Too large. No other local app does this.
Full compatibility table →

Beyond chat — a whole app,
not a runtime.

Already inside, all of it running in the same place.

  • Read your filesPDFs, text, code.
  • See imagesPhotos, screenshots, handwriting.
  • Draw themStable Diffusion, your GPU.
  • Speak insteadTranscribed locally.
  • Run PythonCharts back in the chat.
  • Search the webTwelve sources, off by default.
  • ProjectsStanding instructions, saved personas.
  • Share the engineOne machine serves the network.

For a team: one good machine,
everyone else’s laptop.

The model runs on the computer with the graphics card. Chats and files stay on each person’s own machine. An access key is required, and the host is told what that means for them.

Questions people
ask first.

Short answers. The longer ones are in the docs.

How do I run an LLM locally?
Install Portico, open Models and download one from the built-in list. The inference engine (llama.cpp) ships inside the app, so there is no command line, Python or Docker to set up.
Do I need a graphics card?
No. A GPU makes replies several times faster, but Portico falls back to the processor. Memory matters more: 8 GB of RAM runs a 3B model, 16 GB runs 7–8B, 32 GB runs 32B and up.
Is it really private?
Yes, with one stated exception. The model runs as a process on your machine and the conversation never leaves it. Web search is the only feature that can reach the internet, it is off by default, and the app labels it on screen whenever it is on.
How is it different from LM Studio, Jan, Ollama or GPT4All?
Portico is a finished desktop app rather than a runtime: the engine ships inside it, models are one click, and search, vision, voice, image generation and Python are built in. It also measures your machine and tells you which models actually fit.
Is it free?
Yes — free and open source under the MIT licence. The models are free too. There is no account, no subscription and no message limit.

Try it on your own machine.

Uninstalls like anything else, and takes its models with it.

Download for Windows macOS and Linux

v0.22.0 · 143 MB · SHA-256 on the install page