AI in Castellum
Settings > AI · Castellum
1. What runs where
Castellum's AI features, message summaries, drafted replies, action items pulled from a thread, and Find Similar, run exclusively on a local AI model (Ollama) on your own Mac or LAN. The app enforces a hard egress guard: only a local-network endpoint (loopback or a private LAN address) is accepted, and nothing you haven't explicitly asked about leaves your device. Castellum does not currently offer a cloud-AI option, see §6. Full detail is in the Privacy Policy.
2. Installing Ollama
Ollama is a small, free program that runs AI models locally. To install it:
- Download it from ollama.com and install it like any other Mac app.
- Open Ollama once so it starts its background service (it listens on
127.0.0.1:11434by default). - Pull a model, e.g. from Terminal:
ollama pull llama3.1. Any general-purpose chat model works; smaller models reply faster, larger ones write better summaries and drafts.
If Ollama is already installed and running when you first launch Castellum, the app finds it automatically, see §5.
3. The endpoint
In Settings > AI, the Endpoint field holds the address Castellum talks to, e.g. http://127.0.0.1:11434 for Ollama running on the same Mac. Leave it as the default unless you run Ollama on another machine on your LAN, in which case enter that machine's local network address and port. Castellum's egress guard refuses public/internet endpoints outright, AI features simply won't reach one.
4. Choosing a model
Once the endpoint is reachable, the Model picker lists every model Ollama has installed. Pick one; Castellum remembers your choice and uses it for every AI feature (every AI feature alike). Changing the endpoint clears the picker until you test again, since a different endpoint may serve different models.
5. Test Connection, and first-launch detection
Click Test Connection at any time to confirm Castellum can reach the endpoint and to (re)populate the model list. On a fresh install, if Castellum finds Ollama already answering at 127.0.0.1:11434, it turns AI features on for you automatically and shows a short note under the toggle explaining why, nothing about your mail is sent anywhere by this check. Castellum never auto-enables a non-local endpoint; that only ever happens for Ollama found on your own Mac.
6. Cloud AI, not yet available
Today, AI features in Castellum run exclusively on your local Ollama instance: Settings > AI has no cloud-AI toggle or provider configuration, and no mailbox content ever leaves your device. The Privacy Policy §4 sets out the terms that would apply to an optional cloud-AI mode if Castellum adds one in future, nothing changes unless and until that ships, and it would remain off by default with an explicit opt-in.
7. Where AI jobs show
Every summarise, draft-reply or thread-extraction request becomes a trackable job. Open Window > AI Activity (or the Show AI Activity… button in Settings > AI) to see every job this session, its kind, the message it's working on, its current state, and how long it's been running, with a Cancel button per row and a Clear Finished button once you're done with the completed ones.
8. What the AI does
All of these use your local model only, and need AI switched on in Settings > AI.
- Summarize. Select a message and choose Message > Summarize (
⇧⌘S) for a short summary. - Reply with AI. Choose Message > Reply with AI (it has no default shortcut; you can give it one in key bindings) to have a reply drafted immediately. You edit and send it as normal. There is also a Reply with AI button in the reading pane.
- Thread extractions. Action items, dates and amounts found in a thread appear as small chips in the reading pane. They fill in by themselves when the local model has finished, and nothing shows if there is nothing to extract.
- Find Similar. Right-click a message in the list, or use the Find Similar button in the reading pane, to see other messages that mean something alike. Results open in their own sheet with a similarity indicator, and only when you ask: never while you type a search. This needs an Embedding model chosen in Settings > AI; Castellum builds the index in the background while AI is on.
The jobs for these show up in Window > AI Activity (see §7).