Sotto
Menu
Free · No in-app purchases

Language models that run on your device.

Sotto is a chat app for iPhone, iPad and Mac that runs Apple Intelligence and open-source models locally. There is no account, no server and no analytics — because there is no us for your conversations to reach.

Coming to the App StoreBuild from source

Requires iOS 26.5 or macOS 26.5

Sotto on macOS and iOS. One codebase, one design.

accounts, servers and trackers
0accounts, servers and trackers
inference engines
2inference engines
tools that run on device
25tools that run on device
ways anything reaches the network
4ways anything reaches the network

Two engines

Apple's model, or one you choose

Sotto talks to both, through one interface. You do not need both — if Apple Intelligence is available, there is nothing to download at all.

Apple Intelligence

Apple's on-device foundation model, through the FoundationModels framework. Nothing to download — if your device supports Apple Intelligence and you have turned it on, you can start chatting immediately.

  • No download
  • 4,096-token window
  • Native tool calling

Open models (GGUF)

Any GGUF file, through llama.cpp — Metal on device, CPU in the Simulator. Download one from the built-in catalog or import a file you already have by dropping it on the window.

  • 10 curated models
  • Up to 128K context
  • Metal accelerated

Run them side by side in Compare, or start with the smallest model in the catalog — Qwen2.5 0.5B Instruct, at 398 MB.

Privacy

Four network paths, all started by you

Inference never reaches the network. Everything below is a complete list of what can, and Settings › Privacy shows a running count of the bytes Sotto has actually sent — so you can check this page against the app.

  • Downloading a model

    You start it

    A request to huggingface.co for the file you tapped. The host is pinned in code — a catalog entry pointing anywhere else is refused.

  • The weekly catalog check

    Off by default

    Asks Hugging Face which quantizations exist for the catalog's models. Sends nothing about you.

  • The Google search tool

    Off by default

    Sends only the words the model searched for, under your own API key. You approve the exact query first.

  • An HTTPS tool you write

    You create it

    Goes to the address you configured. Restricted to https, with argument values percent-encoded so a model cannot alter the URL.

Read the privacy policySecurity model

Tools

25 things a model can actually do

Dates and time zones, arithmetic and units, text rewriting, JSON and CSV, statistics, encoding and a search of your own past chats — every one running on the device. 4 ship switched on; the rest are one switch away.

Approval, per tool

Ask every time — which shows a card with the exact URL or command before anything runs — or run automatically. A new tool defaults to asking.

Bounded

At most 4 calls per reply, 20 seconds each, results truncated at 4,000 characters. A model cannot loop.

Why most ship off

Every offered tool spends part of the model’s 4,096-token window before you type anything. Switching on a handful you use works better than switching on everything.

See all 25 tools

What this is not

Models make things up

Sotto runs language models on this device. They invent things, get facts wrong, and can produce text you did not want. Nothing they write is checked by anyone. Treat an answer as a draft and verify anything that matters.

Sotto does not review, filter or fact-check what a model writes — no one does. Model weights you download are third-party works under their own licences, shown next to each model in the catalog. Because it can run open-weight models that are not filtered, Sotto is rated for adults.