AI on your Mac

Everpaper records, transcribes, reads your screen and keeps your journal without any AI. A language model adds the written parts: the running notes and evening entry of your day, decisions and commitments, meeting summaries and the As we go digest, and answers when you ask a question. You can run that model on this Mac, connect one you already use, or use a cloud provider with your consent. You can also go without: in setup, choose Set up later, or Choose advanced options → No AI for now.

What uses AI#

Your options#

Pick one in Settings (⌘,) → AI & summaries → AI connection → AI option.

Nativ is a separate local server app. It is not Everpaper Native. Apple Intelligence is not an option in this beta. On macOS 15 and 26, Everpaper Native doesn't appear in the list.

Everpaper Native (macOS 27)#

On macOS 27, Everpaper can run a model inside the app, with no server to install. Macs with 16 GB of memory or more get Gemma 4 E4B: about a 5.9 GB download, about 8.5 GB on disk (some tools show 7.9 GB), and about 16 GB of free space needed while it installs. Macs with less get Qwen3 4B: about 2 GB to download, 2.3 GB on disk and about 5.5 GB free while it installs.

Setup checks the space first and verifies the download. If the connection drops for a moment, the download carries on where it stopped. Everpaper switches over only after the model has answered a first request, and your current AI keeps working until then. The first request after the engine starts can take up to about 30 seconds. The engine closes after 10 idle minutes.

If you already use Everpaper Native with Qwen on a Mac with 16 GB or more, a one-time card, Gemma 4 E4B writes a better journal, offers the switch. The same offer stays in Settings as Get Gemma 4 E4B. Qwen stays on your Mac until you choose Remove Qwen3 4B on the card, or Remove in Settings.

What setup picks#

In setup's Private summaries step, Everpaper recommends one option for this Mac and downloads nothing until you press Prepare on this Mac. It decides in this order:

1. An AI setup you already have is kept: Everpaper Native, Managed oMLX, LM Studio or Ollama.
2. LM Studio or Ollama, if it's running with a chat model. Everpaper uses it as it is and installs nothing for it.
3. Everpaper Native with Gemma 4 E4B, on macOS 27 with 16 GB of memory or more.
4. Your local app, once it's ready. If LM Studio is installed but not running, or LM Studio or Ollama is running without a chat model, setup says so. Start it or load a model, then press Rescan.
5. Otherwise, Managed oMLX: Gemma 4 E4B on Macs with 24 GB or more, LFM2.5 2.6B below that. On macOS 27 Macs under 16 GB, Everpaper Native with Qwen3 4B instead.

If you chose meetings, Prepare on this Mac also downloads the Everyday meeting models for live captions and speaker labels. The Detailed meeting model isn't part of setup; add it later in Settings → Speech (see Everyday or Detailed).

Choose advanced options shows every route: Everpaper Native (on macOS 27), Managed oMLX, any local servers it found (with Connect), Use my own provider (LM Studio, Ollama, Nativ, a custom endpoint or a cloud provider; finish with Test & Use), and No AI for now.

Managed oMLX#

Everpaper installs the oMLX server (from GitHub) and the model you choose (from Hugging Face), then runs them for you. In setup's advanced options, pick a model and press Install and prepare. The one marked SUGGESTED is the recommendation for this Mac. The choices:

Macs with 48 GB or more also see a few larger models marked as not yet available. They can't be chosen in this beta.

During a Detailed meeting on a Mac with less than 32 GB, Gemma 4 E4B steps aside to leave room for the meeting model. While it does, the As we go digest is skipped, and the summary and journal wait until the meeting ends. If you set Gemma up yourself on a Mac with less than 24 GB, it steps aside during every meeting.

LM Studio, Ollama, Nativ and your own endpoint#

If you already run LM Studio or Ollama, Everpaper uses a chat model it already has; it never downloads into another app. On Macs with 24 GB or more, the suggested model is Gemma 4 E4B: google/gemma-4-e4b in LM Studio, gemma4:e4b in Ollama. OpenAI-Compatible works with any server that speaks the OpenAI API: enter its base URL and model.

Where your text goes#

• No AI set up: nothing is sent.
• Everpaper Native: runs inside the app.
• oMLX, LM Studio, Ollama, Nativ: private while their address is on this Mac, which is the default. Point one at another computer and Everpaper treats it like a cloud provider.
• OpenAI-Compatible: private only while This is a trusted, private endpoint is on. It starts on for an address on this Mac and off for any other. Turn it on only for a server you control: a trusted endpoint receives text without the cloud switch, with only secrets removed.
• Anthropic, OpenAI, Google Gemini, OpenRouter: nothing is sent until you turn on Allow cloud AI to analyze your journal (Settings → AI & summaries → Your memory & summaries). It's off by default and also appears in Settings → Privacy as Allow online journal analysis. Until you turn it on, summaries, Memory and the other AI features wait. Share your profile with cloud AI is a separate switch beneath it.

Secrets such as keys, tokens and card numbers are removed before any AI call. For cloud providers, and endpoints you haven't marked trusted, Everpaper also replaces email addresses, phone numbers and names it detects. When every provider you use is private, Settings says Your AI uses only on-device or trusted private providers — cloud AI is off.

Setting up a model downloads it from Hugging Face, GitHub (oMLX) or Everpaper's download server. See Privacy for everything else.

Change, test or remove#

Open Settings → AI & summaries → AI connection and pick an AI option. Enter a key, address or model if asked; Refresh list reads the models a server offers. Then press Save and Test Connection. Clear disconnects the current AI and removes its key. An asterisk beside the section means nothing is set up or the last test failed.

With Everpaper Native selected, the same place lists its models by name: set one up (Set up or Get, followed by the model's name), Cancel a download, or Remove a model that's installed but not in use. To go through the guided choice again, use Settings → General → Setup & data → Run setup again.

Files and other AI apps#

In Settings → AI & summaries, under Files you include, choose Add Folder to let Memory's Ask answer from your documents. They're indexed on this Mac with an on-device model unless you set an Embedding model provider. If you do, the same rules about where text goes apply to it.

In the same section, under Access from other AI apps, Allow external agents (MCP) (off by default) lets AI apps on this Mac read your people, files, links and connections, with secrets removed (names, email addresses and phone numbers are kept). They never get screenshots, and raw screen text needs a second switch, …including raw screen text (full-text search). What an agent reads is then handled by that app and the AI it uses, which may be a cloud service. Allow cloud AI to analyze your journal doesn't apply to it, so turn this on only for apps you trust. To connect an app, see Command line.

Known limits in this beta#

• Gemma 4 E4B in Everpaper Native hasn't been tested on a real 16 GB Mac yet.
• The first request after the engine starts can take up to about 30 seconds. When the engine restarts to free memory, the next request pauses for a few seconds.
• Evening entries can condense busy days and leave out some meeting details.
• In Memory, the answer check can occasionally remove a correct detail or let an invented name through, so check the sources.
• No new Episodes are written in this beta, so on a new install Memory's Episodes tab stays empty.