Breeze.
Sign inGet started

Station · self-hosted AI chat

The model runs on a machine you can touch.

Breeze is a chat app you host yourself. Your prompts go to a model on your own hardware and stay there. Exactly one feature reaches the internet, and only after you switch it on.

Start chatting Read the source
Inference
localhost:11434
Transcript
your MongoDB
Egress
web search only
Licence
MIT

The boundary

Here is exactly what leaves.

Most private-AI pages ask you to take their word for it. This is the whole map -- every part of Breeze, drawn on the honest side of the line.

Your machine

  • You

    browser

  • Breeze

    Next.js + FastAPI

  • The model

    Ollama · :11434

  • Transcripts

    your MongoDB

Crossed only when you switch it on

The internet

Nothing else is called. Conversation titles are written by the same local model that writes the replies, not by a hosted one.

Read the security guarantees

Instruments

Everything you reach for, and nothing you don't.

The full feature guide

Thinking, shown

Watch the model reason before it answers. The chain of thought collapses out of the way once you have read it.

Web search

Switch it on for a reply that needs the live web. Sources come back cited. Off until you ask.

Images

Drop an image into the thread and the model reads it alongside your text.

Edit and replay

Change a message you already sent and the thread runs again from that point. Regenerate any reply.

Search everything

Find any past chat by title or by something said inside it. Matches are highlighted in place.

Pin what matters

Keep the threads you are living in at the top of the sidebar.

Titles that write themselves

Each thread is named after your first message. Rename or regenerate it from the sidebar.

Export as Markdown

Take any conversation out as a clean Markdown file. It is your transcript.

Installs like an app

Add Breeze to a phone or dock it in the taskbar. It runs from its own window against your own server.

Any Ollama model

Swap models in one config file. If Ollama can pull it, Breeze can talk to it.

Generative UI

Some answers are better drawn.

When a reply is really a trend, a table or a shortlist, Breeze writes the numbers for one and the app draws it. Every example below is a live widget, not a picture -- pick one to see the spec that produced it.

How the grammar works

You asked: How did revenue move this year?

What the model writes

```breeze-ui
{
  "type": "chart",
  "variant": "line",
  "title": "Monthly revenue",
  "unit": "k",
  "x": ["Jan", "Feb", "Mar", "Apr", "May", "Jun"],
  "series": [
    { "name": "2026", "data": [8.2, 8.6, 9.4, 10.1, 11.8, 13.2] },
    { "name": "2025", "data": [7.1, 7.4, 7.9, 8.2, 8.6, 9.0] }
  ]
}
```

What you see

The model fills in a form the app already knows how to draw. Every spec is checked against a fixed list of widgets first, an unrecognised one shows as plain text, and no part of a reply is ever run as code.

Setup

Three steps, in this order.

Full installation guide

  1. 01

    Pull a model

    Point Breeze at any Ollama instance you can reach. Local machine, home server, the box under the desk.

    ollama pull qwen3
  2. 02

    Bring the app up

    Next.js on the front, FastAPI behind it. Both read their config from one env file.

    bun run dev
  3. 03

    Make an account

    Accounts live in your own MongoDB. There is no tenant, no plan, and nobody else on the instance.

    open localhost:3000

Controls

Hands stay on the keys.

  • ⌘KSearch conversations
  • ⌘BToggle the sidebar
  • ⌘⇧OStart a new chat

Take it home

Run it on your own hardware.

Clone it, point it at your model, and the whole thing is yours -- transcripts, accounts, and the machine doing the thinking.

Start chatting Read the docsRead the source