Offline & privacy

iRoof.top keeps your chats on your computer. With a model that runs on your computer too, nothing has to leave it. Offline mode makes that a rule the app enforces, not a habit.

What happens locally

Step	Where it runs	Network
Import, parsing, the style profile	your computer	none
Storage	JSON files and media in the data directory	none
Search in the history, share-link PINs, the media library	your computer	none
Describing photos, stickers, GIFs, videos	the media model, only when you press Describe with AI	one request per item to the model endpoint
Transcribing voice	the media model, when you press Describe with AI on a voice message, or record one	one request to the model endpoint
A reply	the chat model, when you send a message (or a visitor does)	requests to the model endpoint
Compacting a long history	the chat model, as part of a reply, when the history exceeds the budget	one extra request to the model endpoint

Importing never calls a model. The first call happens when you send a message, press Describe or record a voice message. A reply sends the model your exported history (up to the history budget), the style profile and the media descriptions. So where the model runs decides where your chat goes: a cloud endpoint receives it, a model on your computer doesn't.

Run exclusively offline

1. Install Ollama or LM Studio, or run llama.cpp's llama-server.
2. Download a chat model that supports tool calling, and a vision model if you want photos and stickers described.
3. In Settings pick the endpoint and the model, then switch Offline on:

Server	Base URL
Ollama	http://localhost:11434/v1
LM Studio	http://localhost:1234/v1
llama.cpp llama-server	http://localhost:8081/v1 (its default port 8080 is iRoof.top's: start it with llama-server --port 8081)

4. Press Test. It sends a tiny request and shows the model and host that answered. You can switch the network off and chat on.

The desktop app is set up in Settings. It doesn't read your shell or a .env file (see Desktop app), though it does honour ROOFTOP_OFFLINE and OPENAI_* if they are in the environment it was started with. A server reads the environment (and .env):

Homebrew / cargo server

# --offline is the same as ROOFTOP_OFFLINE=1
OPENAI_BASE_URL=http://localhost:11434/v1 OPENAI_MODEL=llama3.1 rooftop --offline

Docker

docker run -p 8080:8080 -v rooftop-data:/data \
  --add-host=host.docker.internal:host-gateway \
  -e ROOFTOP_OFFLINE=1 \
  -e OPENAI_BASE_URL=http://host.docker.internal:11434/v1 -e OPENAI_MODEL=llama3.1 \
  ghcr.io/irooftop/rooftop

In Docker, localhost is the container itself, so use host.docker.internal (the --add-host flag is needed on Linux only). If Ollama refuses the connection, make it listen on all interfaces (OLLAMA_HOST=0.0.0.0).

What Offline mode enforces

	Offline mode
Chat, history summary, media description, voice transcription	refused unless the endpoint is local
MCP servers reached over HTTP	refused unless local, shown as an error in Settings
Internet sharing (<name>.iroof.top)	off: the relay client doesn't start, and the switch is locked
Telegram bot, own or shared	not started
GIFs by URL	the AI can't send a GIF from a URL that isn't local
Wi‑Fi share links	work

A refused request fails before any connection is made, and the message says why: Offline mode: https://openrouter.ai is not a local endpoint. A redirect from a local endpoint to another host is refused too. A small Offline badge sits in the sidebar header while the mode is on. The setting is saved. --offline / ROOFTOP_OFFLINE=1 turn it on regardless of Settings, and Settings can't switch that off.

Local means

localhost, loopback (127.0.0.0/8, ::1), private network addresses (10/8, 172.16/12, 192.168/16, fc00::/7), link-local addresses, and the names *.local and *.internal (such as host.docker.internal). A model on another computer in your home network counts as local, because it never leaves your network. Names are judged by how they are written, they are not looked up.

What Offline mode can't check

• stdio MCP servers are programs on your computer. iRoof.top starts them but can't see what they connect to, so only add ones you trust to stay local.
• Voice transcription uses the chat API (an input_audio part), not the Whisper /audio/transcriptions API. It needs a local server and model that accept audio there. If yours doesn't, the voice message is still saved and sent, only its text is missing.
• Image and speech generation exist only as MCP tools. Offline, use a local MCP server for them.
• The people you share with open your Wi‑Fi links in their own browsers. Offline mode governs this app, not their devices.

Share with friends safely

Link	What happens
Wi‑Fi http://<your IP>:8737/s/<name>	never leaves your local network
Internet https://<name>.iroof.top/s/<name>	the app opens one outgoing WebSocket to the hub, which forwards the visitor's HTTP requests and your app's responses over it

With an Internet link, your export, your chats and the model stay on your computer, and the hub stores none of them. The hub keeps only an account record: your name, a hash of your secret and when it last saw you.

Be clear about what that does not give you. The hub terminates TLS, so, like any reverse proxy, it technically sees the traffic passing through: what the visitor types, the persona's replies, and whatever history the link lets visitors read (Visitors can read the export, default nothing). It forwards only the visitor pages (/s/…, /api/s/…, page assets) and drops cookies in both directions, and your owner UI and settings are never reachable through it. If that is too much trust, use Wi‑Fi links, or run your own relay hub. More on share links.

With Offline mode the iroof.top relay is off. Wi‑Fi links keep working.

The Telegram bot

Answer from Telegram posts to your group:

• new visitor messages (name, link name, text, or a media label with its description or transcript, and a short preview of the message they replied to), and the AI's replies;
• never the exported history. Old messages are not transmitted.

Bots have no end-to-end encryption: the text passes through Telegram in plain text. With the shared bot it also passes through the hub, which keeps only which group belongs to which account. An own bot talks straight to Telegram, without the hub. Offline mode turns both off.

Where your data goes

Data	Where it goes
Exports, chats, settings, API key	the data directory on your computer
Import and style profile	nowhere
History, style profile, your messages	the model endpoint you chose, on each reply
Photos and voice	the media model endpoint, on Describe or when you record a voice message
Visitor messages on a Wi‑Fi link	your computer and your model endpoint
Visitor messages on an Internet link	the hub (in transit), your computer, your model endpoint
Visitor messages with Telegram	additionally Telegram, and the hub for the shared bot
Offline mode	local model endpoint only, Wi‑Fi links only, no Telegram