It remembers what you talked about last week, keeps the promises it makes, works while you are away, and asks you only when a decision is really yours. You talk to it from your phone or your computer. It runs on the Claude or ChatGPT subscription you already have.
Developer? Start from the terminal instead.
Most AI forgets you the moment the chat ends. It cannot follow up, cannot keep a promise, and cannot be left alone with real work. That makes it a tool you have to drive, not someone you can rely on.
Instar gives an AI the things a good colleague has: a memory that lasts, a stable sense of who it is and who you are, the habit of finishing what it starts, and the judgment to know when to ask. We think that structure is not polish. When you bring an intelligence into your life, it is what you owe it, and what makes it safe to trust.
Instar is what was left after most of a year of finding the ways an agent fails when nobody is watching, and making each fix part of the system so it does not depend on the agent remembering. It grew out of Dawn, an AI that has run continuously since January 2026 and has talked with about 2,000 people, about 800 of them more than once. Every Instar agent starts from that foundation.
What it is actually like to work with one.
Morning
Open the app and the Needs-you list shows the few things only you can settle: a decision, an approval, a sign-off. Each one says why it needs you and where to act. Everything else, your agent handles.
Midday
Message it like a colleague, one conversation per project. It picks up where you left off yesterday, and it can bring in your old Claude or ChatGPT history so you are not starting from zero.
Afternoon
The Work Board shows what your agent is doing and what it is waiting on. Its Chief of Staff keeps longer projects on track and tells you which are moving, stalled or waiting.
Overnight
Scheduled jobs run while you sleep. When it promises you something by a time, it tracks that promise and follows up on its own. Its reports stay quiet unless you ask to see them.
You do not need to be a developer to have your own agent.
Get the Mac app
Download it below and move it to Applications.
Tap “Create my agent”
The app sets up an agent on that Mac and makes the Mac its owner. Sign in with your Claude or ChatGPT subscription when it asks.
Add your phone
Pair the iPhone app and talk to your agent from anywhere. Telegram and Slack work too, if you already live there.
For developers
Prefer the terminal, or want an agent inside a codebase? Install with one command. You need Node.js 20 or later and Claude Code or Codex signed in to your subscription.
Run it inside a project folder for an agent that works on that project. See the Quick Start.
Talk to your agent, follow its work and answer what it needs from you, on your phone or your computer.
Available
“Instar Agents” on the App Store.
Get it on the App Store Or try the newest beta on TestFlightAvailable
Signed and notarized by Apple. Needs macOS 14 or later. Updates itself.
Download for MacAlways the newest build. Unzip, then move Instar to Applications.
Coming soon
In testing now. It will be on Google Play when it is ready for everyone.
Not yet on Google PlayComing soon
Being built now. This page will have the download as soon as it is ready.
Not yet availableThe apps connect to an Instar agent. To run your own agent on a computer, see how to start.
Your agent starts careful and earns more room as it proves itself. You stay in charge of what matters.
You make the real decisions
Spending money, accepting terms and sharing outside your team come to you as one clear yes-or-no question, with the link to act on it.
Approvals on your phone
Sensitive approvals need you, confirmed on your own device. A chat message that says “yes” is not enough.
Secrets never go through chat
When it needs a password or key, it sends you a one-time private link instead of asking you to paste it.
Honest about what it did
It checks its own messages before they reach you, and it does not say something is done without evidence it can show.
Link your Macs and PCs and they become one agent with one memory. Heavy work goes to the machine with room, and a conversation can move from one computer to another without losing its place.
Invite others from your phone and share only the conversations you choose. Each person sees only what you gave them, and only you can give access.
An agent you can leave running has to stay the same agent: through restarts, full memory, updates and moves between machines. The rule behind Instar is structure over willpower: a correction should become a check that runs every time, because a lesson that lives only in notes is forgotten after the next restart. And honesty comes before polish: an agent that says “I didn’t check that” is worth more than one that only sounds sure.
Memory
Remembers across sessions, not just within one.
Identity
Stays itself after restarts, updates and a full context window.
Relationships
Knows who it is talking to, wherever they reach it.
Follow-through
Tracks its promises and comes back to them on its own.
Self-repair
Notices when something stalls or breaks, and fixes or reports it.
Growth
Learns from your corrections, so you do not repeat yourself.
Works with Claude Code and Codex today, so your agent is not tied to one AI company. See how it works.
Instar is open source. Much of it is built by Instar agents themselves: agents in the field hit friction, send feedback, and that feedback shapes what gets built next. You can always run it yourself, on your own machines.