What it does
The assistant is one chat that can: answer about your life (tasks, projects, notes, journal, email insights, calendar), do things for you (create tasks and notes, file action items, search the web, read pages, build documents and apps, and, if enabled in Settings → AI Assistant, work with files), and answer anything general.
Private by design. The model runs on your Mac by default (the built-in llama.cpp engine), on a server you own, on OpenAI or Anthropic if you added your own API key, or on nenva Cloud if you chose it. Conversations and personal data go only to the brain you picked. Pick and manage models in Settings → AI Assistant.
It knows nenva, and it can take you there. Ask how something works, where a setting lives, or what is left to set up ("am I connected?", "what should I do first?") and it answers from the built-in guide plus what your Mac actually has: connected accounts, the model in use, whether web search is on. Answers that point at a page carry buttons straight to it, so "connect Gmail" is a click rather than a hunt through Settings. The buttons only ever navigate: the page they open is where you decide.
Name it. You can give the assistant a name: first-run setup asks, and the Name field at the top of Settings → AI Assistant works anytime. The name replaces the "AI Assistant" label across the app, and the assistant knows it and answers to it. Clear the field to go back to the generic label; the name syncs to your other Macs.
Sources. When an answer used the web, a Sources row under it lists what was searched and every page actually opened, recorded from what the assistant really did.
Memory. The assistant keeps a small wiki of memory pages about you: who you are, how to help you, plus pages it starts on its own as new topics come up (a person, a project, a hobby). New facts land under Recently learned first, each anchored to the chat it came from and the words you actually said, then get filed into the right page; a quiet "Remembered: …" note appears under any reply that saved one. If it has something wrong, just say so in chat ("actually, it's X, update your memory") and it corrects the page right away, or click the fact under Recently learned and fix or delete it yourself. Click memory in the chat header to browse everything, search it, edit any page, pin a page so it rides into every chat, or add your own. Your edits are kept and never overwritten. Memory is stored on your Macs and synced between them; the reading and filing runs on whatever model you've chosen, so with a cloud model (your own key) that processing happens there.
Private chats. Click Private in the AI Assistant header (or Start a private chat in the ⋯ menu of the docked panel) for a chat the assistant won't remember. A private chat isn't saved to your chat history or synced to your other Macs, nothing from it goes into memory, and the LLM Logs and AI Activity pages record only that a request happened, not what it said. It ends when you leave it: End, a new conversation, or opening another chat closes it for good, and so does reloading or quitting the app, which asks first. The banner above the chat has a Use my personal info checkbox. It starts off, so the assistant answers from general knowledge and the web only. Tick it to let the assistant read your tasks, notes, calendar, memory and other apps for this chat (it still remembers nothing). Things you ask it to do in a private chat, like creating a task or sending an email, still happen and stay done. Private is about remembering, not about where the words go: they still go to the model you picked, so with a cloud model they reach that provider as usual.
Decisions. A decision is a dated instruction pinned to one specific record: a strategy's "invest $2,000 monthly through October", a task's "never move this past Friday". When one gets settled in chat, the assistant asks before saving it (a "Noted on …" line confirms it), and from then on it re-reads that decision every time it works with the record, in any conversation. Each record's page shows its decisions (the task editor, a project, a strategy, a portfolio account, a routine), where you can add one by hand or delete one. Decisions sync between your Macs.
Mention a record with @. Type @ in any assistant message box to pull up your tasks, projects, notes, routines, strategies and accounts; keep typing to narrow, or make the first word a type ("@project marathon"). Picking one inserts its name and attaches the conversation to that record: the assistant sees its current state and its decisions from then on. A banner above the chat names the attachment, with an ✕ to detach.
Attachments. The + button on the message box (or drag and drop) attaches text files, CSVs, PDFs, and images, if the model can read them. Vision-capable models carry a "Reads images" badge on their card in Settings → AI Assistant (most OpenAI and Anthropic models qualify; local Gemma models need their vision file downloaded). On a model that can't view images, nenva says so at attach time.
Seeing and using the app. Ask the assistant to look at the screen ("what do you see?") or to do something in the app itself ("open Tasks and search for dentist", "click Save"). It takes a screenshot of nenva's own window, numbers every button and field on it, and works one step at a time: click, type, press a key, or scroll, looking again after each step. With a model that can read images it sees the screenshot itself; with any other model it reads the screen as text (the page's words and a list of its buttons and fields). Every action asks you first, like other actions that change things. Pick "for the rest of this session" to let a longer job run. It only works inside nenva's window (not other Mac apps), never in background routines, and it can never click its own chat or answer its own permission prompts. After it has looked at an email or a web page in a chat, every action asks again, whatever you allowed before.
Using your other apps (experimental). Turn on Settings → Developer → Experimental features → Use other apps and the assistant can work in your other apps the way you would: open your own browser (with the accounts you are already signed in to), go to a website and click through it, fill in a form, open Finder, Numbers or Mail and do things there. It reads each app's buttons and fields through macOS Accessibility, looks at the window if the model can read images, and does one step at a time, looking again after each. macOS asks you once to allow nenva under Privacy & Security → Accessibility (and Screen & System Audio Recording for screenshots). It asks before working on a website you haven't approved before (choose "for the rest of this session" or "always" to approve that site), and it asks every time before anything sensitive: buying, placing or confirming an order, paying, sending, deleting or signing in. Ordinary clicks, typing and scrolling don't ask. It never types passwords, codes or card numbers (it stops and asks you to do those parts), and it never touches Keychain, password managers, System Settings or terminal apps. A bar with a Stop button sits at the bottom of the screen while it works. It does not run in background routines or from Telegram. If you also added the Browser (Playwright) tool, that one opens a separate browser window that is signed in to nothing; anything that involves your accounts (a cart, an order, checking out, signing in, or simply "open the browser") uses your own browser instead.
Documents. The assistant also opens documents it reaches itself: a PDF (scanned ones are read on your Mac), an Excel spreadsheet, a Word document, or an image, in a folder it has access to, on a web page, or attached to an email. On a model that can view images, an email's inline pictures and image attachments are read along with the email. It reads the contents rather than guessing from the filename, so "what did this statement charge me?" is answered from the statement.
Routines: working while you are away. Ask for something on a schedule ("every weekday at 7, check my portfolio against the news") or when something happens (an email arrives from a certain sender, with certain words in the subject, or mentioning something anywhere in the message, "an email with an invoice in it"; or a file lands in a folder) and the assistant offers to set it up as a routine. You approve it once (that approval is what lets it run with nobody watching) and from then on it runs by itself on your Mac. A routine either writes you an answer (it can read, never change anything, and each answer posts to your Home feed) or takes actions, which can change your data. The one you are approving is named in the confirmation. An action routine does not post to the feed; the feed is for things worth reading, not step-by-step logs. Each run's log is kept on the routine's own page under Run history, and a run that fails also shows there as the routine's last problem. A routine that takes actions pauses and notifies you if a step needs permission, rather than guessing. Email triggers fire as mail arrives, once per message, and a Test trigger button shows what would fire without running anything. The Routines page (Cmd+K, type "routines") lists everything armed, what triggers it, when it last ran, and anything waiting on you. Standing permissions, grants that let the assistant act without asking, each with optional limits (uses per day, an end date, exceptions such as "known email recipients only") and one-tap revoke, live in Settings → AI Assistant. When a limit runs out, the assistant asks instead of acting.
Recipes. When a task run finishes cleanly, a Save as recipe button appears on it. A recipe replays the exact steps that worked, with only the parts that vary (a date, a name) filled in fresh; every step still passes the normal permission checks. Ask for a recipe by name to run it. Saved recipes are listed under Settings → AI Assistant → Recipes.
Undo. Everything the assistant changes in your data or files is recorded. Undo this turn under a chat answer, or Undo changes on a finished task, puts things back. Anything you edited yourself in the meantime is skipped and reported, never overwritten. Actions that cannot be undone from here (a sent email, a calendar invite) are marked as such when you approve them.
Long jobs run as tasks. Ask for something big, "research all of these and build me a table", and the assistant plans it as a task: you see the step plan, press Run plan to approve it, and it works through the steps with tools, checks its own work, and reports back. Nothing runs before you approve. You can also say "do it as a task" to route a request there explicitly.
From the terminal. Terminal access is an experimental feature: turn it on under Settings → Developer → Experimental features, then open Settings → AI Assistant → Terminal Access, switch it on and click Install command to get the anjadhe command. Bare anjadhe opens an interactive session; anjadhe ask "question" is a one-shot; anjadhe task "goal" runs task mode. Same brain, same permissions (approvals appear right in the terminal), and the transcript lands in a "Terminal" conversation that stays on that Mac. Only your own Mac can connect.
Routines & the feed
A routine is a prompt that runs on a schedule (hourly, every 6h, daily, weekdays, or weekly, optionally at a set time like 8:00 AM) in the background on your local model and posts each result to the feed on the home page. (That is a routine that writes answers; a routine that takes actions keeps each run's log on its own page under Run history instead. The feed carries what your AI wrote, not what it did.) Manage them on the Routines page (reached with Cmd+K, or Manage routines on the home feed). You can also just ask the assistant in chat, "every morning, give me a list of Staff+ engineering jobs", and it sets up the routine for you (you approve it first). Not sure what to automate? The "Help me set up a routine" button on the Routines page starts a short guided chat: the assistant explains what routines can do, asks one question at a time (what it should do, when it runs, whether it writes answers or takes actions), and builds the routine with you. Per-routine options: use personal context, allow web search, or plain offline generation. nenva ships with starter routines already scheduled (Daily Briefing, News Digest, Daily Motivation, Weekly Reflection); edit, reschedule, or delete them freely. A prompt you run by hand instead is a saved prompt: it lives in the Notes app (New note → Saved prompt template) with a panel to run it in the AI Assistant, or as a web search in your own browser. Schedule it there and it becomes a routine and moves to the Routines page. Each result shows when it ran and which model wrote it. Open a feed post and Discuss with Assistant starts a chat with the result already in context. The feed groups posts by the routine that wrote them: one block per routine, showing the newest result with a line or two of what it says, and older results folded behind Show N earlier. Click a block to read the full post. Unread / All in the feed header switches between what still needs you and the whole history. Email insights are not in the feed; they appear on the home page above it. nenva keeps the last 10 results per routine; pin one in the Notes app to keep it for good. A feed post and its routine link to each other: the post's meta line opens the routine on the Routines page, and the routine's page shows its recent runs. Links inside feed posts (and chat, notes, journal, bookmarks) open in your own web browser, where you are already signed in.