Chats

Turn AI on

AI is off until you switch it on. Open Settings (⌘,), choose AI, and turn on AI features under Availability. The rest of the pane appears once it is on, in this order: Connection, the setup for the connection you chose, Behaviour, Finding sources and Advanced.

With AI features off, the chat, its shortcuts and the AI items in MindLake’s menus are hidden or greyed out, and the background features described under Advanced stop too, on your Mac as well as in the cloud: notes are not titled, filed or summarised, and a job already waiting sends nothing. Find sources still works, as a plain search. Your notes and recordings are untouched, and transcription still runs on your Mac.

Choose a connection

Connection is three tiles. Pick one, then set it up in the cards that appear below it.

The setup step when you first open MindLake asks the same question with three cards: On this Mac (Local), Your AI subscription (CLI Agents) and Your own API key (API). You can change it here at any time.

CLI Agents

MindLake looks for agents on your Mac and shows a card for each it can drive: the one in use first, then those installed, then the rest. The arrow beside the heading looks again after you install one.

The last five connect over the Agent Client Protocol. Each runs on the model its own account uses, and when edits are set to be reviewed, MindLake asks you before each change it wants to make. Every agent is also given MindLake’s ranked note search (see MCP server), with nothing to set up.

If Test connection on Claude Code says Claude isn’t signed in, run claude in Terminal, sign in, and test again.

The agent runs the chat. With Claude Code, the background features under Advanced, and the part of Find sources that writes its searches and reasons, run through the same claudeprogram, on the account it is signed in to and billed there, so there is nothing more to set up. Any other agent only runs the chat: the background features go to Claude through Anthropic’s API, with a Claude sign-in or key saved on the Anthropic card under API. Without one, MindLake asks where to run them instead (see Features).

API

Eight provider cards, each with its own key and model. Use on a card with a key, or Set up on one without, makes it the provider the chat runs on and opens it on its key field. Paste the key and press Save; MindLake checks it straight away. Keys are kept in your Mac’s Keychain, and each card links to the page where that provider issues keys.

Test connection asks the provider for the models your key can use. That proves the key without sending a message, so nothing is billed, and it fills the Model menu with the provider’s own list. Until then you can type a model name. Any model you choose must support tool calls, because that is how the assistant reads and edits your notes. The card in use reads “In use” with its model, or “In use · add an API key” until it has a key; Remove deletes a saved key.

If your Anthropic account is signed in with a Claude subscription rather than a key, the Anthropic card says “Signed in with your subscription” and uses that.

Local

Local runs models inside MindLake on your Mac’s graphics chip, with nothing else to install. The This Mac card shows your chip, memory and free disk space, and the model that best fits. Below it, Models on this Mac lists Apple Intelligence and four Gemma 4 models:

• The card for the largest model your Mac’s memory can hold says Recommended for this Mac. One too big for it says “Needs a Mac with N GB of memory” and cannot be downloaded.
• A model you have downloaded but are not using reads “Downloaded · N GB · not in use”, and Use switches to it. Use on the Apple Intelligence card switches back to Apple Intelligence.
• Download fetches the model from Hugging Face and makes it the model in use when it finishes. Cancel stops it; pressing Download again picks up where it stopped. MindLake keeps 2 GB of disk spare and says so if there is not room.
• Models are stored in ~/Library/Application Support/Mindlake/Models, outside your vault, so they never sync. The card in use folds open on Show in Finder, Test speed (how long it takes to load and how many words a second it writes) and Remove.
• A model is loaded when it is needed and let go after five idle minutes, so it only uses memory while it works.

Every Gemma model can run the chat, with the same tools to search, read and edit notes as the cloud routes, and it also runs the background features. Larger models write better and need more memory.

Behaviour

Two settings decide what the assistant may change:

Finding sources

This section picks the search indexes that Find sources asks for reading on a note’s topic. Wikipedia, Papers (OpenAlex), Hacker News and Stack Overflow are built in and need no account. SearXNG (your own server’s address), Brave Search, Tavily, Exa, Kagi and Google Programmable Search add general web results with your own key, and each shows numbered steps for getting one. Find sources also uses a model to write its searches, pick what is worth reading and say why each page fits. It runs where the background features do (see Features); with no model to use, or with AI features off, it is a plain search. See Sources.

Advanced

Most people never open this. Click Advanced to unfold it.

Features

Each feature has a menu of the models it can run on: under On this Mac, your Gemma model (when one is in use) and Apple Intelligence; under Cloud, Claude Opus, Sonnet or Haiku, or, when API uses another provider, that provider with the model picked on its card. With the Local connection only the first group is offered. With CLI Agents or API only the second is, except for Transcript summaries: their menu offers both, and shows where summaries really run, which is on your Mac unless you pick a cloud model.

Transcript summaries use Apple Intelligence (on macOS 26) unless a Gemma model is in use, and go to the cloud only if you pick a cloud model for them. With the Local connection, every other background feature runs on your Mac too. With CLI Agents or API, where the cloud ones go depends on the connection:

With CLI Agents, or with API on another provider, a note under the list says the same. When they would go to Anthropic’s API and no Claude sign-in or key is saved, they can’t run, and a section headed Without a Claude account asks where to run them instead, with a button for each choice your Mac has: your Gemma model (when one is in use), Apple Intelligence (on macOS 26), or another provider whose key is saved. Once you pick, it says where they run, and Change picks again or stops them. If what you picked is no longer set up on this Mac (its key removed, or its model deleted), it counts as nothing picked and the section asks again. With CLI Agents, the pill over the chat shows an amber dot until then (see Open the chat). Save a Claude sign-in or key later and they go back to Claude.

Understanding

The search index behind the assistant’s note search. Find notes by meaning is off by default; turn it on and the assistant also finds notes that say the same thing in other words. It uses the sentence model built into macOS, so nothing is uploaded and there is nothing to download. The section shows Notes indexed, how many are Understood while the first pass runs, and a Rebuild index button. The index lives only on your Mac, is a cache that changes nothing in your vault, and never includes locked notes.

Open the chat

The chat lives in the drawer on the right. Open it in any of these ways:

• The drawer button at the right of a screen’s header.
• ⌘/ (View ▸ Show AI Panel), which also closes it.
• ⇧⌘A (File ▸ Ask Claude), which always opens it on what you are looking at.
• Ask the AI from the command bar, or /ask there, which opens the chat with your question drafted. See Search and the command bar.
• Send to Chat on a task, which quotes it into your message without sending it.
• Talk about it on a task, which starts a new chat already listening (see Talk about a task).

Press ⏎ to send and ⇧⏎ for a new line. While the assistant works, small pills show each step, and Send becomes Stop, which ends the reply but keeps the conversation. The new chat button in the drawer’s header starts again. The drawer’s title is a menu that also switches it to Calendar, Sources, Preview or Terminal. Two fingers up or down over the title step through them without opening the menu, skipping any with nothing to show.

A small pill above the conversation names what is answering, with its logo: your Gemma model, marked On this Mac, when the chat runs on your Mac; the model picked on the provider’s card with API; the Claude model with Claude Code; or, with any other agent, the agent itself, which runs on its own model. It changes as soon as you switch any of them. Click it to see what the chat runs on and what the background features run on, under Chat and Background jobs; AI Settings… opens Settings ▸ AI. An amber dot on the pill means the background features have nowhere to run: with an agent other than Claude Code, no Claude sign-in or key is saved and nothing has been picked to run them instead. The detail names what is stuck, and Choose a Model… opens Settings ▸ AI ▸ Advanced at Without a Claude account (see Features).

Terminal is a real login shell on your Mac. It starts in the folder of the note or canvas you have open, or the folder you are looking at, or else your vault’s folder. It is separate from the assistant: it works whichever agent you use, and the assistant does not use it.

What the assistant can see

The scope menu in the drawer’s header offers Whole vault, plus This folder, This note, This canvas or This list when one is on screen. It follows what you are looking at until your first message; changing it after that starts a new chat.

• The open note. In This note, the assistant starts out knowing the note’s folder, date, tags and people, and which notes sit closest to it. Work you ask for, such as a summary, goes into the note unless you say otherwise. It can also read the note’s sources that are ticked for the chat.
• The rest of your vault. It searches MindLake’s index of your vault, ranked best first, and can narrow by tag, person, folder or date. With Find notes by meaning on, it also finds notes that say the same thing in other words.
• Tasks and contacts. Your open tasks (on by default) and your Apple Contacts (off by default), as set under Advanced.
• Memory. Lasting preferences it learns go in memory.txt in your vault, which every chat reads. Settings ▸ Brain ▸ Open memory file shows it.
• Not locked notes. Locked notes are never available to the assistant, even while they are unlocked on screen.

Let it edit your notes

With Let the AI take actions on, the assistant can create notes and edit them. When a reply changes a note, a card under it names the note and counts the lines added and removed; it folds away after a few seconds and reopens when clicked. A note open on screen flashes as it changes. With Review each change first, each change waits on its card until you press Approve or Decline.

Checks for lost content

MindLake compares every AI edit with the note before it, looking for a removed heading, a dropped task, a transcript cut by more than a tenth, or more than a tenth of the words gone. With an API provider or a Gemma model, an edit like that is refused before it is written. With a CLI agent it goes through, since you may have asked for the cut, and an alert says what was removed.

Edit history

Before taking in any AI edit, MindLake keeps the note as it was. The last 20 versions of each note sit in ~/Library/Application Support/Mindlake/EditHistory, one folder per note, outside your vault so they never sync. Canvases are kept the same way. There is no restore button yet: open the folder in Finder to copy an older version back.

Clean Up and Retag & Link

A note’s ⋯ menu has two one-step jobs, shown while AI features is on:

Each touches only that note. If the result would lose content, the note is put back and you are told why; otherwise Undo is on offer for 10 seconds. Both always run through Claude Code, whatever connection you chose, so Claude Code must be installed and signed in on your Mac.

The Chats section

Every conversation is saved as a Markdown file in your vault’s Chats folder, so it syncs with the vault but never shows up as a note. Chats in the sidebar (⌃5) lists them newest first, grouped into Today, Yesterday, This Week and Earlier. Each opens in the drawer, where a new message carries it on. A reopened chat shows its messages, not the step pills or edit cards.

• Titles. A row is titled by your first message. A chat about something gets a prefix in a quieter tone: Note:, Task:, Calendar:, Canvas:, List: or Folder:. A chat that began with a quoted task is titled by the task.
• Length bar. Five steps beside each row show how far it went, counted in your messages: none filled is a stub (no real answer), one a one-shot, two one follow-up, three 3 to 4 messages, four 5 to 7, five 8 or more. The bar turns orange when the chat ended without a reply.
• Under the title. The latest reply, and the notes the chat read or edited.

The filter button in the header narrows the list; the search box matches titles and replies.

Select and delete

One conversation goes to the Trash straight away. More than one asks first, with Move to Trash. Either way, ⌘Z (Edit ▸ Undo) brings them back. Right-click a row, or use its ⋯ button, for Open and Delete.

What leaves your Mac

Nothing is sent to an AI provider until you connect one. After that it depends on the route: