01
Yours
The thinking happens on your computer. What it knows about you is one file you can read or delete.
Local-first personal AI
An assistant that lives on your computer instead of someone else's.
It runs on your own hardware and remembers you for years. Type to it, or talk to it out loud. Your calendar, your email, your shell, your lights. No account, no subscription, nothing leaving the machine unless you ask.
See how it works When can I run it?
$ cp env.example .env
$ make setup
$ make start-all
# → http://localhost:8080, and it's yours from herewhat have I got on tomorrow, and is it going to rain
calendar_list_events search_web memory_recall
Three things. Standup at 9:30, the Kew Gardens site visit 11:00 to 13:00, then dinner with Charlotte at 19:30.
Rain from about 10 until mid-afternoon, so the site visit is the wet one. You walked it last time and complained about it, so I'd take the car.
book the car and tell Charlotte I might be 15 late
Waiting for you
email_send → charlotte@example.com
“Hi Charlotte, I'm Luna, Marcus' assistant. He may be about 15 minutes late…”
- 100%on your machine, works offline
- Dozensof tools, before you add any
- Anywhereyou already are
- 1file holds its memory of you
What it is
Four things make it different from a chatbot with a local model.
Most assistants forget you between sessions, or remember you on someone else's server. This one keeps what it learns in a file on your disk.
02
Remembers
Facts, and a record of every conversation you've had. It searches by what you meant, and reaches back months.
03
Reachable
Web, voice, Telegram, Slack, your editor. One memory behind all of them.
04
Acts
Your files, terminal, email, calendar, house, and code. On a schedule, if you like.
How it works
The pipeline every message goes through.
Whichever channel you use, your message runs the same steps. That's why the one on Telegram knows what the one in your browser knows.
Doors in
- Web UI voice or text
- Telegram text, voice notes, photos
- Slack DMs and mentions
- Your editor while you're coding
- Robot mode hold space, talk
The core, every time
1
Listen, if you spoke rather than typed
2
Remember who you are and what you've said before
3
Work out which of its tools this needs
4
Go and do it, however many steps that takes
5
Check with you before anything it can't undo
6
Answer, out loud if you spoke
On step 3: it works out which of its tools your question needs. It speaks MCP too, so any tool server you point it at joins the same pool.
All on your machine
- The model the part that thinks
- A second, smaller one doing the background work
- Its memory of you a file you own
- Its tools plus any MCP server you add
- Its ears and voice
It runs in reverse too. A scheduled task takes the same route out, and finds you wherever you are: your browser if it's open, Telegram if it isn't.
Memory
Memory that outlives the session.
When a conversation ends it writes the thing up in the background: what happened, how it went, anything about you that came up. Weeks later it can find that again.
Declarative
Things it knows
Facts, preferences, and reminders, weighted by how much they matter. “What's my pet called?” finds “dog is named Max”.
Episodic
Things you did together
What each conversation was about, written up without you asking. This is what it searches when you refer back to something you discussed.
Working
The thread you're on
What you're talking about now. It reaches back for the rest only when it's needed.
…so Jodi reckons the timeline slips to March unless we cut the migration.
episode saved topics: timeline, migration · tone: tense
21 days · 40 conversations · one restart later
what did Jodi say about the timeline again?
conversation_recall
That she'd expect March if the migration stays in scope. You weren't thrilled about it. Want me to dig out what we said about cutting it?
Ask what someone said three weeks ago and it'll tell you.
Everywhere
Reachable wherever you are, with one memory behind it.
Each one starts only if you've set it up. The rest never load.
The web UI
Type, or hold to talk. Anything it builds for you opens in a window you can drag, resize, and use.
Ask for a tic-tac-toe game and you get one you can play.
Out loud
Hold the button and talk. It answers out loud, and it writes differently when it knows you're listening, so you get something that sounds spoken rather than a paragraph read at you.
The listening and the speaking both happen on your machine, so nothing you say is sent anywhere to be transcribed. The voice is yours to pick.
Telegram
Text it, send a voice note or a photo, get replies in kind. Lock it to your own account and nobody else gets in.
In groups it only replies when mentioned.
Slack
In your workspace as a bot, answering DMs and mentions. It'll post on your behalf too, asking first any time it would be speaking as you.
Your editor
It stands in for the cloud assistant your editor expects, so Zed, Cursor, or your own script can use it with nothing to adapt. Same memory, same tools.
Robot mode
A full-screen robot face with eyes that follow you. Hold space and talk.
No tools, no memory, nothing it shouldn't say. Children find this considerably more interesting than you will.
Its own identity
It writes as itself, on your behalf.
Every email, message, and invite it sends goes out as the assistant and says so in the first line. It doesn't get to bend that.
It looks you up before writing, refers to you in the third person, and signs off as itself. Talking to you directly it drops all of that.
The person receiving it knows who wrote it and that you asked for it, which a ghostwriter can't give you.
Yours to name. The name, the voice, and the personality are all yours to set. Luna is only the default.
L
Luna
on behalf of Marcus
| To | john@example.com |
|---|---|
| Subject | Coffee next week? |
Hi John,
I'm Luna, Marcus's assistant.
Marcus would love to grab a coffee next week and pick up where you left off in June. He's free Tuesday afternoon or Thursday morning. Shall I put something in?
Luna
PA to Marcus
Proactive
It starts the conversation.
Tell it when, in plain English. It writes itself a scheduled task, and the result finds you wherever you are.
- “Every weekday at 7, brief me on my calendar and the news.”
- “Check for new email every ten minutes, only tell me if it matters.”
- “Tuesdays, draft me something to post about AI. Don't send it.”
What that's like to live with
- You describe the job and the timing in your own words. No syntax, no form.
- Shut your laptop and missed jobs run once when it wakes.
- It finds you in your browser, or on Telegram if you've gone out.
What it can do
Plenty of tools, and it picks them itself.
About half are there from the first run. The rest reach into your accounts, and wait until you connect them one at a time.
offline works with the network off online needs the internet, nothing else needs a key needs your account
Files
offlineRead, write, edit, search, and organise, inside a workspace you set. Anything destructive asks first.
Your machine
offlineRuns commands on your machine, from a list you choose. Read-only, developer, or more.
Memory
offlineRemembers, recalls, and forgets on request. You can read or delete anything it holds.
Images
offlineMakes pictures on your own machine. With a model that can see, it reads the ones you send it too.
Artefacts
offlineWrites a document or a small working app, then opens it in a window you can use.
Skills
offlineTeach it a procedure in plain markdown and it follows it. No rebuild needed.
The web
onlineSearches and reads pages, and drives a real browser when a site needs clicking through. No account needed for any of it.
Email and calendar
needs a keyRead, search, send, and file your mail. Move events across every calendar you own, and find gaps everyone shares.
Slack
needs a keyReads your DMs and mentions, searches history, sends messages. As a bot, or as you.
Your house
needs a keyLights, plugs, thermostats, locks, sensors. “Is the heating on?” becomes answerable.
Code and issues
needs a keyYour pull requests, issues, and tickets. It'll review a branch and tell you what it thinks.
Pictures
It draws, and it can see.
Ask for an image and you get one, made on your own hardware. No credits, no queue, no monthly allowance.
You describe roughly what you want and it writes the detailed prompt itself. No need to have learned how to talk to an image model.
Nothing about the request leaves your machine. Your family, your house, an idea you haven't shown anyone: none of it becomes someone else's training data.
It works the other way round too. Send it a photo and it'll tell you what's in it, read the text off a screenshot, or use it as the reference for the next thing it draws. That one depends on the model you're running: some can see, some can't, and it's a line of config either way.
It can email one to someone or send it to your phone from the same conversation.
I need an image for a landing page I'm putting together about you
Your desk, late on. Say the word and I'll try something else.
Made by it, on the machine it runs on.
Built to keep
What you'd notice after a month.
The point of running your own is that it keeps working on your terms. A few of the choices behind that you'll notice.
You're not tied to one model
It runs on whatever you already have: LM Studio, Ollama, llama.cpp, or something you built yourself. Swap in a better model next month and nothing else changes. The assistant you've built up shouldn't be thrown out every time a better model lands.
The flip side is that how well it does any of this tracks the model you give it. A small one on a modest laptop picks its tools less sharply and writes less well than something larger. The wiring is identical either way, so upgrading the model upgrades everything at once.
Run it however suits your machine
Containers, or straight on your machine, or split between the two with the model on bare metal where it's quickest. Nothing forces one shape on you.
Remembering doesn't slow you down
Writing up conversations, tidying memories, and pulling out facts all happen in the background, on a second smaller model. You never wait for the bookkeeping.
Very little to go wrong
It leans on what the language and the browser already do rather than stacking up frameworks. Fewer packages to patch, less to break, less to read if you want to look inside. Something you rely on daily should be boring to maintain.
It stays quick on ordinary hardware
Most of the work went into keeping it sharp on a laptop rather than a rack. The result is a personal assistant rather than a science project that wants a datacentre.
Trust and control
It has your shell and your inbox. These are the limits on it.
Destructive actions need approval
Deleting a file, running a command, sending an email as you: it stops, shows you what it's about to do, and waits. Same from your phone.
It can only run what you've allowed
Its access is a list you choose, limited to folders you name. Anything off that list can't run, whatever it decides it wants. Start with almost none.
It doesn't trust what it reads
Emails and web pages can carry text written to trick an assistant. Anything from outside is marked untrusted and treated as information rather than instructions. If a page tries to give it orders, you hear about it.
What it can reach
Thinking and remembering happen on your machine, and it works with the wifi off. Asking about your inbox does mean talking to your inbox, but nothing is connected until you connect it. Haven't given it your email? Then no amount of asking gets it there.
Your data is a file
Everything it knows about you is one file. Copy it, read it, take it to your next computer, or delete it. Nobody to ask.
Availability
Getting it ready to share.
This runs every day on my own machine. Getting it ready to run on everyone else's requires a little more work.
Nothing to sign up to yet. If you want to hear when there is, I'm easy enough to find.
Whether it'll run on your machine
| Machine | Apple Silicon Mac or Linux. |
|---|---|
| Memory | 16 GB minimum. Ideally 32 GB depending on your local models. |
| Disk | About 30 GB all in. Most of that is model weights. |
| Setup |
Docker is quickest. You can run the pieces directly if you'd rather. I run llama.cpp and
stable-diffusion.cpp on the host, and everything else in docker.
|
| Sign-up | None. Connect your own accounts for email, calendar, Slack, issue tracking, and the rest. |
Why run your own
Every conversation adds to what it knows about you, and that record is a file on a disk in your house. It gets more useful the longer you run it.