You get an actual file
Documents, tables, decks: it writes them into a file you open and edit, with nothing to copy out of a chat window. The first time you use one of these formats it takes a moment to prepare.
Word · Excel · PPT · PDF
Unichr is an AI agent that lives on your computer.
Ask it to draft a memo or tidy up a spreadsheet. It finishes the job, and what you get back is ready to use.
THE FIRST SCREEN
Pick a line and it lands in the composer verbatim, cursor at the end, ready for you to keep typing.
Where do you want to start?
Pick a line · it lands verbatim
I'll paste in scattered content — make it one Excel sheet, you pick the columns.
This is the first screen you see after opening it.
SHIPPING TODAY
All six run in the current macOS beta.
Documents, tables, decks: it writes them into a file you open and edit, with nothing to copy out of a chat window. The first time you use one of these formats it takes a moment to prepare.
Word · Excel · PPT · PDF
It works inside files already on your machine, organised by project. You don't have to move your work somewhere else to use it.
File tree · Projects
Anything outbound or irreversible waits in one place for your OK. Everything reversible, it finishes on its own.
To-do inbox
Finished work collects in Results, so you can go back to it later. If a step went wrong, return to that instruction and run it again.
Results · Checkpoints
It remembers the edits you make and the formats you prefer, all on your own machine, so next time takes less work.
Memory Graph
Hand it something you do every day and set a time. It runs by itself and delivers the result wherever you asked.
Cron jobs
Which model it connects to is up to you. See Install below.
THE BET
Every job you hand it teaches it a little more about how you work. That stays on your own machine and is never uploaded, so each time goes a bit smoother.
What you build up stays yours,and moves with you.
NON-NEGOTIABLE
It runs through these five before every action. When it isn't sure, it stops and asks.
CONFIRM
Anything that spends money, goes out, or can't be undone waits for your confirmation.
NO FAKING
If it can't do something, it says so. It won't tell you a job is done when it isn't.
EXPLAIN
It can tell you why it did something the way it did. And you can look up every step it took, or undo it.
SAFE
It won't put your accounts, files, sessions or public standing at a risk it can't control.
LOCAL
Nothing on your machine leaves it without your explicit say-so.
NOT YET
Today it is a generalist. It doesn't know your line of work, and it won't decide what counts as finished. These are on the way. None of them exists today.
It would have a defined role, know what to watch, and know what makes a deliverable good enough. People who know the work build those, you install one, and you can take it out if it doesn't fit. That is what we are working on next.
These depend on the layer above. There is no entry point for them in the product today, and we won't pretend otherwise.
The “nothing to prepare, just install and go” route isn't open yet.
We're getting the macOS build right first. The other two haven't been released.
GET STARTED
Once installed, connect a model. Two ways in, and you can switch at any time.
If you already have a ChatGPT subscription, one click opens a browser sign-in and there are no keys to dig up. It's the only account you can connect directly on first run.
Request an invite →macOS only for now (Apple Silicon), by invitation.
OpenRouter
One key, hundreds of models. The default choice if you have no strong preference.
Regional providers work too
DeepSeek, Kimi, GLM and MiniMax, alongside Anthropic, OpenAI and Gemini.
A model running on your own machine
Point it at something local and nothing leaves your computer at all.
Unichr's own free allowance isn't open yet. For now you'll need an existing ChatGPT subscription or an API key of your own. You can also skip that step and add one later in Settings.
FAQ
For now, yes. Either sign in with a ChatGPT subscription, or paste in a key of your own (OpenRouter, DeepSeek, Kimi, GLM, MiniMax, Anthropic, OpenAI, Gemini, or a model running locally). Our own free allowance isn't open yet.
Your files, your history and everything it builds up for you stay on your own machine. But whatever you hand it to work on goes to the model provider you chose, because that is where the reasoning happens. If that matters to you, point it at a model running locally and nothing leaves your computer.
No. Anything outbound or irreversible stops and waits for your OK. Every step also has a way back, so if it went wrong you can return and run it again.
Not yet. Today it's a generalist. You give it a job and it does it, but it doesn't know the conventions of your field and won't decide what counts as finished. Teaching it your trade is what we're working on next.
Nothing during the beta, and there is no payment path in the product at all. The model side runs on your own subscription or key, and we don't mark it up.
The macOS beta is invitation-only. Leave an email and we'll open them one by one.
Request an invite →