Cope

Start with a problem, or without one

One issue, one conversation, never ended

You pause it by leaving and resume it by coming back. After a break, Cope starts with a short recap and the actions that came due.

Follow-through, not advice

A daily reminder tells you what's due. Coming back, Cope goes through what was due, asks what actually happened, compares it with what you expected, and says what that means for the next step. From a simulated run in the showcase, a week after the plan was set:

The same simulated person as the panel further down, two visits on. Read the whole issue in the showcase.

Independent voices, and the choice handed back

One capable model, asked a hard question, finds the safe middle and gives it to you in one voice. Asked for both sides, it writes both already knowing how it wants them to end. Cope keeps its voices apart instead, so that when they agree, the agreement means more, and when they split, the split is real. Council apps keep models apart too; Cope's voices each hold a stance of their own, and a split comes back to you by kind. It doesn't stage a fight: most questions get an answer. This is the bet the whole thing rests on.

Some issues belong to more than one person

A couple's money, a company two founders run, a parent that siblings look after, a coach's work with a client. Invite someone into an issue (Invite, on its More tab) and you work it through together, live, each on your own device. Up to three of you.

• Cope mediates. It answers once everyone has spoken or passed, and treats what each person says as that person's account, not the facts. It takes no side and doesn't tell one of you what the other really means. A disagreement you can't settle yet is written down as one, each view with its owner's name.
• Everyone's yes. Everyone gets the same suggestions and answers them on their own device. A change goes into the document only if everyone says yes; anyone can leave it out or ask to discuss it.
• Everyone keeps the record. Each person has a full copy of the issue. It moves on only while everyone is there, so the copies never disagree; between sessions it's read-only for all of you, and each of you is reminded of what's due.
• Nothing hidden. Whoever you send the link to sees everything in the issue, the whole conversation and every revision, and Cope tells you so before you send it. No one has a private thread.

Keep ChatGPT. Bring the issues that run for weeks here.

For thinking one question through today, a frontier chat is as good as Cope or better, faster, and already paid for. In our own test, a plain chat on the same model came out even with Cope on a single problem. Cope is built for what a chat isn't: the issue that is still there next month.

• A record you approve, not a memory written about you. A chat assistant writes its own notes about you as you talk; you find out what it kept afterwards. Cope changes nothing without your yes, keeps every change as a revision you can undo, and keeps what it knows about you as entries you approved one by one.
• A plan that is checked, not a conversation that ends. Each step gets a date and an expected result, and when you come back Cope asks what actually happened and what that says about the next step. A chat carries what it remembers of you, not a plan it comes back to.
• Voices with a stance of their own, not one voice. Asked for both sides, one model writes both already knowing how it wants them to end. Council apps run several models apart, with a chair's answer as the result. Cope's consultants each hold a stance of their own, answer blind on two models by default, and where they split, the split comes back to you by kind: a fact to find out, a prediction to check, a call that is yours.
• Nothing of it on our servers, account or not. A chat assistant runs on its maker's servers; what you say is stored there, with or without an account. With your own key, Cope keeps your issues on your device and nothing you say reaches us; sign in only if you want us to hold them.
• An issue two people work through, with everyone's yes. In a chat the assistant answers whoever typed last, and nothing is held in common. Cope keeps one issue that is both of yours, each on your own device, and it changes only when all of you say yes.

Each does one part of it well

By kind of product, with one checked example each. What a row says is about the kind of thing a product is, not a feature it lacks this month.

You don't have to trust us with anything

Cope asks you about the things you'd least want on a server: your marriage, your boss, your money, what the Serpent told you to do. So it's built so that they need not be on ours. No account, no email, no sign-up: open the app with your own API key, and nothing you say or write reaches us. We built this, and your data isn't what we're after.

How far you can take it. Three steps, each with one party fewer in your conversation.

1. Signed in, on Cope's models. Your conversations pass through our server on their way to the providers, and stay in your account if you keep your data there.
2. Your own key. Only the model provider you chose sees them, under its own API terms, and the app names it before the first call. Your data stays on your device.
3. Your own model, on your own machine. Nobody. In the terminal client not a word leaves it; the browser app makes one request to our host, to load, and no other. We've written up the setup, and we're measuring which models carry a Cope conversation. The same works with a model server run by someone you trust, for their circle: they see what the model sees. The guide for both is coming.

Voice is the exception: speech goes to the browser's maker or to the voice provider you choose, so the last step is typed.

What reaches us on every step, exactly. Loading the app from copeai.app is a request to our host like any website's; the terminal client makes none. The daily reminder, if you turn it on, hands our relay times and ciphertext it can't read. Inviting someone into an issue hands our rooms your devices' addresses, and everything between you is encrypted end to end. None of these carries a word you wrote.

If you'd rather we hold your data, you can, and this is exactly what that means. Privacy or convenience is your choice, and we recommend neither. Two choices, made apart: whose keys, and where your data lives.

Cope comes with its own test bench

In the simulator, AI users bringing real kinds of problems use the real app over several visits, and a judge model scores each run blind, on a written rubric. A second test plays the same problem against the same model with nothing of Cope around it, a plain chat with one frontier model, and a judge compares the two conversations, in both orders so the order doesn't decide. New parts are tested the same way: the check-up also meets a simulated person whose money is in order, to catch it inventing problems, and a shared issue is played with a simulated person on each side.

• We publish the results, good or bad. Claims about helping people with their lives are easy to make; we want to show that Cope works, and keep showing it, including when it doesn't. So far: the first comparison with a plain chat came out even, and About you's first three simulated runs showed no sign yet that it helps, and a wrong entry in the list was checked rather than followed.
• Anyone can run it. The simulator runs in the browser app, on your own keys or your credits, with your own simulated people and scenarios beside ours and your own cast, so you can test your own problems, your own changes, or check our claims.
• It's a research question too. Does a room of opinionated voices, run by a moderator, help a person more than one capable model on its own? Work on multi-agent systems asks something like it, mostly on tasks with right answers; Cope asks it about personal problems, where "better" is judged against a rubric. The answer could go either way, and we'll publish it either way.

Five of those simulated people are open to read: their issues as they left them in the app, and the same runs in the simulator with their judgements.

Write a consultant. Test it. Publish it.

Cope's cast is a starting point, not the last word. The method is plain text: a consultant is a page that says what it holds and whose thinking it stands on, and a playbook is a page on how a kind of problem is worked through. Someone who has lived a co-founder dispute, bought a first home, raised a child with ADHD or run a small restaurant knows things no general assistant will say with conviction. They can write that voice. A coach can write playbooks for their clients; a community can build a room of its own, a Stoic room or a founders' room. Cope's own consultants and playbooks stay as they are: you can choose them, not rewrite them.

• Write it in the app. Your own consultants and playbooks sit beside Cope's, switched on and off like the rest, and Cope calls them in when their lens fits.
• Test it before you trust it. Run it in the simulator against our simulated people or your own, read the panels, and judge the runs blind on the same rubric we use.
• Publish it. Signed in, publish it under a handle you choose to the community page, where anyone reads the prompts and installs them with one click, with the same plain warning every pack carries. We read every listing before it goes public, and a listing shows how many have installed it, no ratings.
• What it earns. Today an accepted pack earns its author Cope credits. A share for authors of what their work earns is for later, once Cope itself earns.

Try it on us

Cope is new, and what we need most is people using it on a real problem and telling us how it went. Write us a few lines: what you'd like to try, and a line about you. We read every request, usually within a few days, and while places last we add $5 of Cope credits to your account to start. Use it for real, tell us what you think, and there's more.

• Your feedback is what you write to us. We never read it out of your conversations.
• Nothing public is asked. No reviews, posts or ratings: we want to hear what didn't work as much as what did.
• Found a bug? Search the bug tracker first, and add yours there if it's new: the first report of a bug, with the steps to reproduce it, earns credits. No account? Write to info@copeai.app and we'll file it for you.
• Other help earns credits too: consultants and playbooks others can import; judging simulator runs blind, to check our judge against people; and trying what no stranger has tried yet: signing in from an Outlook address, reminders and calls on a phone, a shared issue with someone.

The honest limits

• Not a therapist or a crisis service. Cope has no safety layer today: no crisis detection, no hotline prompts, and the consultants speak without softening. A check-up raises hard subjects on its own (your inner life, a partner, your family).
• Not built to keep you talking. Each issue is meant to end. The assistant will call only on the schedule you set or when something in your plan is due, and it looks for problems only when you ask for a check-up.
• What it costs. Each conversation is real model calls at the providers' prices, which depend on the models you choose; the cost is shown per issue. Cope's models through CopeRouter cost their makers' prices plus a small margin, paid with Cope credits; with your own keys you pay only the providers.
• Where it could fail. You may tap yes without reading. A record of your decisions can pin you to a self you've outgrown. Voices from one model can sound like one voice in masks. Talking a problem through can become a way of not acting on it. A check-up can talk you into a problem you don't have, or measure your life against a standard that isn't yours. In a shared issue, one person can press another into a yes, and Cope's refusal to take sides is no protection against that. These are the risks of the very features that make Cope worth using, and we name them so you can watch for them.