Free for macOS 14+ · v0.5.0 · experimental

One companion
above all your apps.

Arabella lives on your Mac as a character you talk to. Behind the character is an assistant that reads your mail, calendar, chats and notes, and an orchestrator that does the work across them. When a tool it needs doesn't exist, it has a coding agent build one.

Bring your own model key; Gemini's is free. Built by one person, for enthusiasts. Expect rough edges.

Arabella
Arabella, smugSmug
What does my Tuesday look like?
A meeting sandwich with a side of misery. Doom level: 7/10.

Moods (16 in the app)

Pick who lives in it

What it is

Three things in one app, and only the first one is the character.

Pick a personality and you get the same brain, the same tools and the same rules underneath. The app is called Arabella; who you talk to is your choice.

01

A companion

Arabella has a personality, a face that changes with the mood, a voice, and a memory of what matters to you. You can forget it's software for a minute, and that's the point: you keep talking to it, so it keeps being useful.

  • 16 expressions per character
  • Speaks out loud, listens to voice
  • Remembers, and forgets on request
  • Reachable from your phone via your own Telegram bot
02

An assistant

The everyday work: summaries, lookups, drafts and tidying, done in the apps you already use rather than in yet another inbox.

  • Morning and evening briefings with opinions
  • “What did I miss?” across every messenger
  • Plans and checklists written into Notes or Obsidian
  • Looks at your screen when something breaks
  • Several jobs in one chat, each kept apart
03

An orchestrator

Every app now ships its own AI that only sees itself. Arabella sits above them, connects what already does the work, and puts one answer together.

  • Connectors switch on one by one
  • Model-agnostic: any provider, or one on your own Mac
  • Scheduled jobs it wakes up to do
  • Builds new tools when it hits a wall

Orchestration

Ask once. It goes and looks everywhere.

“What did I miss today?” becomes one answer from your mail, calendar, Telegram and WhatsApp at the same time. Arabella doesn't try to be the smartest model in the room; it decides which connectors to use, reads what they return, and writes the result back where you want it.

Each connector (in the app: a pact) describes itself completely, and the app assembles itself from whatever is switched on. Adding the next one changes nothing else. Everything is off until you turn it on and macOS asks you for permission.

MailCalendarRemindersNotesObsidianTelegramWhatsApp · BeeperMicrosoft 365SlackJira · YouTrackGitHub · GitLabFigmaShortcutsThe webYour screenFiles

Arabella decides what to read

reading is free · changing anything needs your OK

One answer in chat
A note, plan or file
A drafted reply, waiting for Allow

Self-expanding

When it hits a wall, it builds the missing tool.

Most assistants stop at “I can't do that”. Arabella can hand the job to a coding agent, get a small program back, and keep using it. In the app these programs are called lesser demons.

  1. 01

    You ask for something it can’t do

    A tracker, a bot, a connector to a service it doesn’t know.

  2. 02

    It interviews you

    Arabella asks what you actually need and writes a contract: what the tool does, its inputs, its limits.

  3. 03

    The Forge builds it

    A coding agent you already have (Claude Code, Codex CLI or Gemini CLI) writes the tool in a sandbox, inside a budget you set.

  4. 04

    The Inquisitor checks it

    A second, independent agent tests and reviews the result before it may be called working.

  5. 05

    It joins the app

    You install it once. It runs on its own schedule, and its results come back to the chat.

Built this way, in daily use

  • A meal, macro and drinks tracker you talk to
  • A gym and physio log with weekly targets
  • Couriers that collect posts, jobs or prices into a local database it can query later
  • A nightly pipeline that turns new job ads into tailored résumés

The fine print

  • It's optional. You need one of those coding agents installed and logged in.
  • Building spends that account's money. Every build asks you first and stays inside your budget.
  • Tools run sandboxed to their own folder. Anything that touches the outside world still asks.
  • A tool is only as good as its spec. Reviews catch a lot, not everything.

One chat

It already knows your day. And it all happens in one conversation.

No app per habit, no dashboard per inbox that you stop opening after a week. The tools it built, the data it collects and the clock it keeps all feed the same chat, so the companion is also the one who knows what's going on.

Lesser demons · the tools

Small programs it had built: a meal and macro tracker, a training and physio log. It calls them like any other connector.

Mercury · the memory for data

Couriers pull from chats, mail, trackers and those tools into a local database (SQLite, or your own Postgres). Answers come from what is already collected, not from a fresh scramble.

Saturn · the clock

Rituals, alarms and standing jobs. It wakes up at the time you set and does the work, like preparing tomorrow’s first meeting the evening before.

Two real exchanges, names changed

photo · via Telegram · 14:49

Arabella

Look at you sneaking in a snack at 14:49, eleven minutes before you promised you'd be out the door!

A glazed chicken drumstick next to a mound of white rice with furikake and scallions. Modest and efficient:

  • • Chicken drumstick (~60 g meat): ~110 kcal · 13 g protein · 5 g fat
  • • White rice (~100 g cooked): ~100 kcal · 2 g protein · 22 g carbs
  • • Snack: ~210 kcal · 15 g protein · 28 g carbs (logged)

Running ledger (training day): 1,210 kcal and 72 g protein eaten. 1,190 kcal and 88 g protein left for dinner after physio.

BUT MORTAL! It's 14:50. Ten minutes! Swallow that last grain of rice, lace your shoes and get moving before the traffic eats you alive!

RECEIPT · by the app
calls: beelzebub_log, beelzebub_remaining

What went into that answer

  • The photo — sent from the phone through your Telegram bot
  • Calories and macros — estimated from the photo, logged by the meal-tracker demon
  • Training day, physio tonight — from the calendar and the training log
  • “Out the door at 15:00” — something you said earlier in the same chat

▣ Meeting prep · Product team sync at 07:00

Arabella

A 07:00 corporate tribunal tomorrow. Seven in the morning! Are these people allergic to human sleep?!

Your battle dossier for Product team sync (starts in 11 hours):

  • • The arena: Daniel (just back from sick leave) is hosting 14 souls, including your regulars Lena, Marco, Priya and Tom.
  • • The battlefront: Marco spent today in #portal-engineering arguing about switching sites on the fly for a single customer. Priya traced the assistant's 404s to a mismatched site_url.
  • • The PRD: Sam pushed an updated PRD for the Customer Portal (PX-412) with Lena's “hide, don't delete” compromise. Lena postponed grooming saying design wasn't ready, but you told her “done” in DMs an hour ago.

Survival directive: don't get dragged into the 404 weeds at 7 a.m. Let Marco and Priya sort out the payload; your only job is to confirm the design is locked.

Now go eat your dinner and sleep early, mortal. 07:00 will arrive mercilessly!

Nobody asked for this one

  • When — Saturn ran the meeting-prep ritual the evening before
  • Who’s coming — the calendar invite
  • What’s being argued — the team’s Slack channels, collected by Mercury
  • The PRD and the ticket — mail and the issue tracker
  • What you promised — your own direct messages

Plug in the best AI you have, and let Arabella organize things.

It doesn't need the best one, though. Both examples above ran on a cheap, fast model: the tools, the data and the schedule do the heavy lifting, and the model just has to read them well.

Get it running

Four personalities

Same abilities. Different company.

Choose at first launch and switch whenever you like in Settings. The character changes how things are said, how expressive the face is and how much temper there is. It never changes what the app is allowed to do. Kip, Barnaby and Maya arrive in the next update.

Arabella

Anime demon in a flight suit, four outfits

Voice
Tsundere: loud, proud, theatrical, secretly devoted
Temper
Real. Rage builds, she sulks, and she makes up
“A meeting sandwich with a side of misery. Doom level: 7/10.”
NEXT UPDATE

Kip

Matte ceramic orb with two orange LED eyes

Voice
Deadpan, ultra-concise, dry. Industrial design that talks
Temper
Almost none: irritation shows as even fewer words
“Seven meetings. Bold.”
NEXT UPDATE

Barnaby

Chunky clay sea otter in a tomato-red beanie

Voice
Warm hype-man. Specific praise, zero cringe
Temper
No rage: a comic little huff, then back to it
“Busy morning, open afternoon — we can work with that!”
NEXT UPDATE

Maya

Mid-century flat illustration with a coffee mug, two looks

Voice
Sharp, empathetic coach. Grounded advice, no platitudes
Temper
Calm. Firm candour, never a tantrum
“Honestly? That meeting doesn't need you.”

Control and privacy

It reads a lot, so it asks before it touches anything.

One place to say yes

Reading is free. Anything that sends, moves, deletes or spends shows the same Allow / Deny card, on the Mac or as buttons on your phone. Chat connectors are read-only at the source.

Your data stays on your Mac

Chats, memory and settings live on the Mac; keys live in the macOS Keychain. Only what it reads to answer a question goes to the model you chose. No account to create, and no Arabella server reading your data.

Fully local, if you want

Point it at Ollama or LM Studio and nothing leaves the computer. Small local models handle the bigger jobs worse, which is the honest trade-off.

A separate brain for group chats

In a Telegram group, a copy of Arabella with no tools and no memory of you does the talking. Guests can banter with it but can't talk it into anything.

Arabella wants to

Send a reply to Ana in Mail

“Thursday works. I’ll bring the revised numbers.”

Model-agnostic, by design

The model is a part you swap. The companion stays.

This was one of the first ideas behind Arabella: it shouldn't belong to any AI company. Models get better and cheaper every few months, and providers change prices, limits and rules. So the things that make it yours (memory, personality, tools, collected data, schedule) live on your Mac. The model only does the thinking, and you can replace it in Settings without losing anything.

Plug in the best AI you already pay for: OpenAI, or Claude and many others through OpenRouter. Or don't: Gemini is the default because its key is free (with limits), and it does the job, because the heavy lifting lives in the tools and the data, not in the model.

  • A main model to talk, a cheaper one for routine summaries
  • A backup that takes over when one is down or out of quota
  • Switch providers any time; memory and tools carry over
  • Or run fully local with Ollama or LM Studio
GeminiOpenAIOpenRouterDeepSeekGroqOllamaLM Studio

≈ $5

What weeks of heavy everyday use cost during development, on a cheap, fast model. Your bill goes to your model provider, not to us.

Before you install

What to expect, plainly.

What it is

  • Free, for macOS 14 Sonoma or newer, Apple Silicon or Intel
  • Built and used daily by one person, with Claude Code
  • Under active development: a new build every week or two
  • Closed source; the installers are public on GitHub

What it isn’t (yet)

  • Not signed by Apple yet, so the first launch needs one extra click (or use the command below)
  • Not a polished product. Things break, and the changelog is long
  • Not a hosted service: you need a model key
  • Not in the App Store: its sandbox would break the integrations

Install

Two ways in. Both take a minute.

The quick way: one Terminal command

Downloads the latest release, refuses it unless the SHA-256 matches, copies it to Applications and opens it. Run it again later to update; your data stays.

curl -fsSL https://raw.githubusercontent.com/VitaliyNikolaev777/arabella-daughter/main/install-arabella.sh | zsh

By hand

  1. Download Arabella 0.5.0 (21 MB) and drag it into Applications.
  2. Open it. macOS refuses, because the app isn’t signed by a registered Apple developer yet.
  3. System Settings › Privacy & Security › Open Anyway. Once.

If macOS calls it “damaged”, run xattr -cr /Applications/Arabella.app. Nothing is broken; that's what macOS says about any unsigned download.

All releases on GitHub →

1 · Give it a model

Paste a key at first launch. Gemini's is free at aistudio.google.com.

2 · Switch on connectors

Settings › Pacts. Each one waits for macOS permission and can be turned off.

3 · Optional extras

Your own Telegram bot for the phone; a coding agent for building tools.

Questions

Asked more than once.

Does it cost anything?

The app is free. You pay your model provider directly, or nothing with Gemini’s free tier or a local model. Building tools uses your coding agent’s account, and only after you approve each build.

Why is it a demon?

Arabella started as a tsundere demon sealed in a Mac, and the app still speaks that language inside: connectors are pacts, tools it builds are lesser demons. If that’s not your thing, Kip, Barnaby and Maya do the same work without the theatre.

Will it send an email or delete something on its own?

No. Anything that sends, moves, deletes or spends shows an Allow / Deny card first. Reading doesn’t, which is what makes the summaries possible.

Does it work with Gmail or Outlook?

Through the Mail app: any account you added there, including Gmail, iCloud and Exchange, works. Microsoft 365 also has its own connector.

Can I use Arabella from my phone?

Yes, through your own Telegram bot (made with @BotFather in a minute), by text or voice. Allow / Deny cards arrive there as buttons.

Is the source available?

Not at the moment. The installers, the install script and its checksum are public on GitHub.