Using your own model

AI summaries and the security scan for your end-to-end encrypted mail run in your browser, against an AI model you choose. Your mail is sent from your browser to that model and nowhere else. Joinery never sees it. Whoever runs the model does, so choose one you trust: a service you hold a key for, or a model on a computer of your own.

Setting it up

Choose your model. Open Email settings and find Your AI model. When the site runs a model of its own, Use this site's model is one click. Otherwise, or to use another, choose Enter a different model and give the address your model's service gives you (for example https://api.fireworks.ai/inference/v1), its key (empty if it needs none) and the model's name. You confirm it is you with your passkey, because this decides where your mail may be sent. The start of the address, up to the host name, is kept with your account; the rest, the key and the model name stay in this browser, so on another computer you enter those again in the AI panel on your mail.
Open your mail. The AI panel checks the model on its own. When it can be reached, all that stays is one line, "Your model … is reachable", with Change for the details. If your browser asks whether this site may reach your computer or network, allow it. When something stopped the check, the panel says what and the fix. Behind Change, Test sends the security scan's real instructions with a made-up sample message, for when you change something or want to check the context.
Turn the AI on for the mailbox. In the same panel, turn on Email triage for one-line summaries, Email security scan for the danger check, or both. On an end-to-end encrypted mailbox each card says Runs on your device while this mailbox is open.

What you will see

A one-line summary under each new message in the list, as your model finishes it, while the mailbox is open and your vault unlocked.
The security scan's verdict and its red flags in the message, marked judged by the model's name.
Summarize and Scan now under an opened message, for one message at a time, with the answer shown at once.

The work runs in one browser tab at a time and judges up to 200 messages per visit, so a large mailbox is done over a few visits. When your model cannot be reached, the panel says so in one line and the mailbox keeps working without AI. Nothing errors in the reader.

The site's own model, with one click

When the site you use runs its own model on hardware on its network, Email settings offers it first, and the AI panel offers it too: Use this site's model, naming the model and the address. One click sets the address and the model for you. You still confirm with your passkey, because that confirmation is what allows the mail page to send your mail there. The offer says plainly who runs that machine: if you are not the site's operator, the operator could see mail sent to it, so choose it only if you trust them with your mail.

A service you hold a key for

Fireworks, OpenAI and any host that speaks the OpenAI chat format (vLLM, LM Studio and others) all work. Paste the base address the service gives you for chat completions, without /chat/completions on the end.

Your domain's AI setting may keep mail on your own machines. If it does, the panel says the address you chose is not allowed for that mailbox; the domain's administrator can loosen the setting on the domain page.

A model on a computer of your own

Ollama on the computer you are using works with three settings on your side.

Let this site call it. Ollama answers only sites named in its OLLAMA_ORIGINS setting. Add your Joinery site's address (for example https://mail.example.com) to that setting on the computer running Ollama and restart Ollama. The address to paste in Email settings is http://localhost:11434/v1; the key stays empty; the model name is the one you pulled, such as llama3.1:8b.
Allow your browser once. Chrome asks before a web page may reach a program on your computer or your network. Allow it once for your site and it remembers.
A model that thinks is fine. The page asks it not to reason unless the recipe's thinking level says so, the same way the site's own AI does, so its answer is not spent on reasoning.
Give the model enough room. A message and its instructions come to about 3,600 tokens. Ollama's chat address takes no context setting from the page, so set the model's context on your side, for example with a Modelfile line PARAMETER num_ctx 8192. A context that is too small fails with a message Test shows word for word.

A computer elsewhere on your network works the same way with its address in place of localhost. A plain http:// address is accepted only for localhost and for private or Tailscale network addresses written as numbers; a name needs https://. tailscale serve gives an Ollama an https:// address on your Tailscale network, and that site address must be on Ollama's origins line too. Safari and Firefox may refuse a plain http:// call from a secure page; the https:// route avoids that.

What the check says, and what to do

Reachable, with the model's name and address: nothing to do.
Your browser asks first or your browser blocked it: allow the site to reach your computer in the browser's prompt or site settings.
Your model refused this site: add the site to Ollama's origins line and restart Ollama.
Wrong key: check the key with your service.
No such model: check the model's name; on Ollama, pull it first.
Context too small: the model's own message. Raise the model's context.

Where things are kept

The host name of your model is kept with your account and named in the mail page's own security policy, so the page can send mail to that host and to no other. Changing it asks you to confirm it is you again. Removing it never asks: it can only narrow where your mail goes.
The rest of the address, the key and the model name are kept in each browser you enter them in. They never reach Joinery.
The site grades your model by its name. When it grades below what the security scan asks for, the panel says so, and the verdicts may be less reliable.