AI providers and models

Every AI feature in Nixt Mail uses one destination that you choose. This page covers each kind of destination, where keys are kept, how to name a model, how to check that it answers, and every message you can see while setting it up.

If the AI row shows a sentence instead of choices, AI is not available in your copy, edition or licence. See AI overview.

Open the AI settings

1. Open Settings with ⌘, (Ctrl+, on Windows and Linux), or choose Settings… from the Nixt Mail menu on macOS or the File menu on Windows and Linux.
2. In the System group, choose AI and services.
3. Under Where AI and services run, find the row named AI.

The row shows a sentence describing the current choice, then the places AI can run:

Your edition may not offer every choice. When it holds one back, a line under the choices says which choices your edition does run AI on, and names an edition that includes the one it holds back.

The rest of the form depends on what you choose, and appears under the choices.

Use a model on this computer

A model running on this computer keeps everything on this computer. Nixt Mail does not include a model runtime: you install one yourself, and Nixt Mail finds it.

1. In the AI row, choose Your own server.
2. Wait for Nixt Mail to finish Looking for a model on this machine….
3. Do one of the following:
   • If a runtime is found, click one of the model names listed under it. Nixt Mail fills in the address and the model name for you.
   • If nothing is found, click Set one up on this machine and follow the guided setup.
   • Or type the address and model name yourself. See Enter the address yourself.
4. Click Check it answers. See Check that it answers.

Runtimes Nixt Mail looks for

Nixt Mail looks only on this computer, at the port each runtime uses out of the box. It never looks anywhere else on your network.

What you may see under the choices:

The install suggestion depends on your system. On macOS it is:

brew install ollama

On Windows:

winget install Ollama.Ollama

On Linux:

curl -fsSL https://ollama.com/install.sh | sh

Installing a runtime is up to you. These are the commands Nixt Mail shows; check the runtime’s own site before running anything.

When a runtime is running with no model, the suggestion is:

Set up a model with guided setup

Set one up on this machine opens a window that reads how much memory and free disk space this computer has, looks for a runtime, and walks you through getting a model. Each step says what it is about to do before it does it.

1. Click Set one up on this machine.
2. The window shows Looking at this machine… and then a line describing the computer, such as its memory, cores, processor type and free space.
3. What happens next depends on what it finds:

Before anything is downloaded, the window shows a sentence naming the exact command it will run, the size of the download, and where it comes from — the runtime, not Nixt Mail — and confirms that nothing of yours is sent to obtain it. The download does not start until you click Download.

The models offered are:

A model is offered only when its memory figure fits in the computer’s total memory minus 4 GB, and its download plus 2 GB fits in the free disk space.

Other messages from guided setup:

A runtime Nixt Mail starts keeps running after you close the window. Nixt Mail does not stop it.

Enter the address yourself

1. In the AI row, choose Your own server.
2. Type the address of the model server into the address field.
3. Type the model name into the model field, exactly as the server knows it.

The address must be a complete web address:

• On this computer, http://localhost:11434, http://127.0.0.1:1234 and similar addresses are allowed.
• Anywhere else, the address must start with https://.
• If you give only a host and port, Nixt Mail adds /v1/ to it.

The server must accept requests in the OpenAI chat completions format, with streaming.

Use your own server

Use this for a model server run by you or your organisation somewhere other than this computer.

1. In the AI row, choose Your own server.
2. Type the server’s address, for example https://ai.example.com.
3. Type the model name, exactly as the server knows it.
4. Click Check it answers.

For a server that requires an API key, choose Your AI provider, then OpenAI-compatible, and enter the server’s address and your key there.

Messages under the address field:

Use your AI provider

Use this when you have your own account and API key with a provider. Requests go straight from this computer to the provider, under your account and your agreement with them. Nixt Mail is not in the path and never sees the reply.

1. In the AI row, choose Your AI provider.
2. Under Whose model, under your own account with them., choose the provider. Until you choose, the form says Choose one. Nixt Mail will not pick a company to send your mail to.
3. For Azure OpenAI or OpenAI-compatible, type the address of your deployment or service.
4. Type the model name, exactly as your provider spells it.
5. Type your API key into the Your API key field, then click Save key or press Return.
6. Click Check it answers.

The address for Azure OpenAI and OpenAI-compatible follows the same rules as your own server: https://, or http:// only on this computer. An OpenAI-compatible address on this computer keeps everything on this computer, and the row says so.

To reach Amazon Bedrock through a gateway that accepts the OpenAI format, choose OpenAI-compatible.

Where your key is kept

• The key goes straight into this device’s keychain. It is not written to settings, not put in the body of a request, and never shown again.
• On macOS, it is kept for this device only: it is not synchronised through iCloud and does not move to another Mac with a backup.
• It is kept against the destination: the provider and its address. If you switch to another provider, Nixt Mail looks for that provider’s key instead. If you switch back, the earlier key is used again.
• Once a key is stored, the field reads A key is stored. Type a new one to replace it. and the line under it reads Held in this device’s keychain, for this provider only. It is never put in a request body and never shown again.

If the key cannot be stored, a message says This device has no keychain service running, so the key could not be stored. On Linux, this means no Secret Service, such as GNOME Keyring or KWallet, is running. Start one and save the key again.

To remove stored keys, use Settings › Privacy and security › Erase everything on this device. That removes everything Nixt Mail keeps on this device, not only keys.

Choose a model

The model name is required. Nixt Mail never picks a model for you, because a small model on a laptop and a large hosted model are not the same, and only you know which your mail deserves.

• Type the name exactly as your runtime or provider spells it.
• When the name includes a size, such as 8b or 14b, a line under the field estimates the memory it needs: At 4-bit this wants roughly N GB of free memory. On a machine with less, a smaller model answers rather than swaps.
• Changing the model name does not end the permissions you granted. Those are tied to where requests go, not to which model answers.

Check that it answers

Check it answers sends one fixed line, “Reply with the single word: ready”, and allows a reply of one token. It contains nothing from your mail, so you can run it before granting any permission.

Changing the placement, the provider, or the host name in the address puts the line back to Not checked yet.

What changing the destination does

While you hold permissions for the current destination, a line in the form warns: Changing this ends N permission(s) you granted for the current destination. You will be asked again.

Changing the placement, the provider, or the host name in the address ends every permission that named the old destination, and the next request asks again. So does choosing Off. See AI permissions.

Timeouts

A request fails if the destination does not start answering within 30 seconds, or stops sending for 60 seconds in the middle of an answer. Anything that had already arrived in the Assistant is kept.

Error messages

These messages appear in the Assistant, in a snack bar under the reply box or compose window, or beside Check it answers. host is the destination’s host name.

If a provider repeats your key in an error message, Nixt Mail replaces it with [your key, hidden] before showing the message.