Getting Portico running.

Portico runs on Windows, macOS and Linux. Windows has a ready-made installer; on macOS and Linux you build it yourself for now, which takes about five minutes and three commands. Everything below is the whole process, including the parts that look alarming but aren't.

Straight answer about the other two: the Windows build is the one that is tested on every release. The macOS and Linux instructions are written from how the project is put together, and no one has yet run them start to finish on those systems. They should work — Electron and llama.cpp both build there routinely — but you would be the first to find out. If you try, say what happened, whether it worked or not.

What you need.

Memory

8 GB minimum, 16 GB comfortable. This is the real limit on which models you can run — more than the processor or the graphics card.

Disk

About 400 MB for the app, plus whatever the models take. 2 GB gets you a good first model; collections grow quickly.

Graphics

Optional. A Vulkan-capable card (NVIDIA, AMD, Intel) or Apple Silicon makes replies several times faster. Without one it falls back to the processor and still works.

Internet

Only to download the app and your models. After that Portico works with the connection off — unless you switch web search on yourself.

Nothing else is required. The inference engine and the speech engine ship inside the app; there is no Python, no Docker, no account and no API key.

Windows 10 or 11, 64-bit.

The quickest route — one installer, no build step.

  1. Download the installer

    About 126 MB.

    Try Portico

  2. Run it, and expect a warning

    Windows will say “Windows protected your PC” and call the publisher unknown. That is not a sign anything is wrong — it appears for every application that hasn't paid for a code-signing certificate, which costs a few hundred pounds a year.

    Click More info, then Run anyway.

  3. Choose where it goes

    The default is fine. Portico installs for your user account only, so it doesn't ask for administrator rights.

  4. Open it and pick a model

    Portico starts with no model — it is the record player, not the records. Go to Models → Discover and download one. See choosing your first model below.

macOS 11 or newer.

Apple Silicon (M1 and later) or Intel. On Apple Silicon the model runs on Metal, which is genuinely fast. There is no prebuilt .dmg yet, so you build one.

  1. Install the prerequisites

    Node.js (version 20 or newer), and Apple's command line tools:

    xcode-select --install

  2. Get the source and its dependencies

    git clone <the Portico repository>
cd PriStudio
npm install

  3. Add the inference engine

    The engine is built per platform, so it isn't in the repository. Download the latest release from llama.cpp and unzip llama-server together with every .dylib beside it into:

    resources/llama-mac-arm64/     # Apple Silicon — asset: bin-macos-arm64
resources/llama-mac-x64/       # Intel        — asset: bin-macos-x64

    Flat, no subfolders.

  4. Build it

    npm run dist:mac

    The .dmg and a .zip appear in dist/. Drag the app to Applications.

  5. Open it the first time

    The app isn't notarised by Apple, so double-clicking gives “Portico cannot be opened because the developer cannot be verified.” Right-click the app and choose Open, then confirm in the dialog. macOS remembers your choice — you only do this once.

Any modern 64-bit distribution.

Tested against Debian and Ubuntu conventions; the AppImage should run anywhere reasonably current.

  1. Install the prerequisites

    Node.js 20 or newer, and — for graphics acceleration — a Vulkan driver:

    # Debian / Ubuntu
sudo apt install nodejs npm mesa-vulkan-drivers

# Fedora
sudo dnf install nodejs vulkan-loader mesa-vulkan-drivers

# Arch
sudo pacman -S nodejs npm vulkan-icd-loader

    On an NVIDIA card, the proprietary driver already provides Vulkan. You can skip this entirely and run on the processor — slower, but it works.

  2. Get the source and its dependencies

    git clone <the Portico repository>
cd PriStudio
npm install

  3. Add the inference engine

    From the llama.cpp releases, take the bin-ubuntu-vulkan-x64 asset and unzip llama-server plus its .so files into:

    resources/llama-linux-x64/

    If you would rather have no driver dependency at all, use the plain bin-ubuntu-x64 asset instead — processor only.

  4. Build it

    npm run dist:linux

    You get both an .AppImage and a .deb in dist/.

  5. Run it

    # portable, no install
chmod +x Portico-*.AppImage
./Portico-*.AppImage

# or install the package
sudo dpkg -i portico_*.deb

Your first few minutes.

The same on every platform.

1 · Test your graphics

Settings → Test graphics cards. Portico times a real reply on each card in your machine and keeps the fastest.

This is worth doing even if you think you know the answer. On laptops the discrete card is often powered down by the driver, which can make the built-in graphics several times faster — we measured exactly that while building this.

2 · Download a model

Models → Discover. Pick one that fits your memory:

  • 8 GB RAM — a 3B model, about 2 GB
  • 16 GB RAM — a 7B or 8B model, about 5 GB
  • 32 GB RAM — a 14B model and upward

A model bigger than your memory will still load, but pages from disk on every word and becomes far too slow to use.

3 · Start typing

That's it. The strip under the message box shows which model is loaded, how hard it is working, and how full the conversation memory is.

Web search is off until you switch it on, and the app says so on screen whenever it is running.

What works where.

Chat and everything around it works on all three systems. Two features are Windows-only for now, because the projects behind those engines don't publish Mac or Linux builds — they have to be compiled separately.

FeatureWindowsmacOSLinux
Chat with local modelsYesYesYes
Web searchYesYesYes
File attachmentsYesYesYes
Reading images (vision)YesYesYes
Projects and assistantsYesYesYes
Shared engine over your networkYesYesYes
Voice inputYesBuild it yourselfBuild it yourself
Image generationYesBuild it yourselfBuild it yourself

Portico runs perfectly well without those two — they simply show as unavailable rather than breaking anything.

When it misbehaves.

Windows blocks the installer

Expected. Click More info → Run anyway. The app isn't code-signed, which is a cost rather than a safety judgement. If your browser refuses the download outright, choose Keep in the downloads bar.

macOS says the developer cannot be verified

Also expected. Right-click the app and choose Open instead of double-clicking, then confirm. Only needed once.

A model won't load

Almost always memory. Lower Context size in Settings, or use a smaller model. Portico tells you when the engine ran out of memory rather than failing silently.

Replies are very slow

Run Settings → Test graphics cards. If the fastest result is only a few words per second, the model is probably running on the processor — check that your Vulkan driver is installed, or pick a smaller model.

It forgot the start of the conversation

The context window filled up. Click the usage ring in the strip below the message box to see where it went, and raise Context size — at the cost of memory on your graphics card.

Nobody on the network can reach my shared engine

Windows Firewall usually asks to allow Portico the first time — say yes, on the private network. If it was refused, allow it again in Windows Defender Firewall settings. Also check both machines really are on the same network, and that the address and access key match what the host shows.

Something crashed

Settings → About → Copy diagnostics puts versions, settings and the last few log lines on your clipboard. Paste that into a bug report. Nothing is sent anywhere unless you send it.

Removing it.

Windows

Settings → Apps → Portico → Uninstall. Or run Uninstall Portico.exe from the install folder.

macOS

Drag Portico from Applications to the Bin.

Linux

Delete the AppImage, or sudo apt remove portico if you installed the package.

Your chats, projects and downloaded models live in your user data folder and are not removed by uninstalling — so reinstalling keeps everything. Delete that folder by hand if you want a clean slate. Models are the big items; everything else is tiny.

Try it on your own machine.

Free and open source. Works with any open model you like. Nothing you type leaves your computer — unless you switch on web search.

Download for Windows
Version 0.21.0 143 MB Windows 10/11 · 64-bit

On a Mac or Linux? Build it in three commands →