How it works
Four things carry the whole product: covens, agents, threads, and memory. The rest is setup you do once.
A coven is one world.
Make one for each client, employer, or venture. A coven holds its own agents, its own threads, and its own memory, and you choose which of your AI accounts and apps it may use.
Strix is the coven you start with. It’s the one for your own life, and it’s free. What AICoven remembers about you lives there, and your other covens don’t read it.
Northwind
No other coven’s agents read this coven’s memory.
Agents do the work.
An agent is a named role with instructions: what it’s for and how it should go about it. You choose who a message goes to, and the agent that answered is always named.
Each agent runs on one AI account and one model. If that model is unavailable, the turn stops and tells you which model failed and how to fix it. An agent never switches to another model by itself.
Can you check the numbers in the board pack against the ledger?
To Ledger · in Northwind
A thread is one piece of work.
You ask, and you get an answer. While the agent works there is one line saying what it’s doing. The steps it took are behind Details if you want them, and out of the way if you don’t.
As a thread grows it keeps its own account of what was decided and what is still open, so you can leave it for a week and pick it up.
Ledger
Two figures don’t match. Page 4 shows Q4 revenue as £412,000; the ledger says £409,300. The gap is one invoice dated 2 January.
Worked 2 minutes · 3 steps
You were last here on Tuesday. Still open: whether the January invoice belongs in Q4.
Memory is what should be true before you say a word.
Threads are not memory. Memory is the short page of things every agent in a coven should already know.
- An agent asks“Remember this for the coven?” with the line it wants to keep.
- You approve or denyWhat you approve joins the coven’s Memory page, where you can edit or forget it later.
- Every agent in the coven reads itWhichever AI model it runs on. No other coven does.
Handing work off.
Give an agent a job and carry on. Some steps need your OK first: sending a message in your name, for one. The request appears beside what you’re writing, waits in Activity if you’re elsewhere, and can be answered from its notification.
How much an agent may do in a connected app is your choice per coven: read only, or read and make changes.
Ask an agent to repeat a job, every day or on the days of the week you choose. It writes each run into the thread you asked in.
Contract renewal in Northwind
Send this to Dana in Slack, as you?
“The redline is on its way. Fee and notice are agreed; payment terms are still open.”
Where a message goes.
The same path every time, whichever AI account or agent does the work.
- Your deviceYou send a message. The app holds a key made on your device and kept in iCloud Keychain, and sends it with the request.
- AICovenOur servers build the turn: the agent’s instructions, the coven’s memory, the recent thread. Content is readable in memory while the request is handled.
- The AI you choseThe turn goes to your own AI account, to an agent you host, or to your Mac. Their terms cover what they do with it.
- Back, and stored encryptedThe answer returns to your devices. Your message and the answer are stored encrypted, and your devices decrypt them to show them.
Under the hood.
For the technical reader. None of this is needed to use the app.
- Account keys
- Each provider key is encrypted with its own data key using AES-256-GCM, and that data key is encrypted with a master key.
- Message content
- Encrypted with AES-GCM under a 256-bit key generated on your device. Three kinds of content use other keys: repeating jobs (a key our servers hold for 30 days at a time), shared covens (the coven’s own key), and assets our servers must be able to read (a key per asset). Some things are stored readable: the titles of threads, tasks, memories, and assets; agent names and instructions; chat attachments; voice transcripts; and a repeating job’s title and prompt. A repeating job set up without a key, which the app marks “Not encrypted”, also stores what its runs write as readable text.
- Routes
- Cloud AI accounts are called from our API. Hermes and OpenClaw run where you host them. OpenAI-compatible servers get chat only, not agent runs. Agents and tools on your Mac are reached through the Mac app while it is running.
- Workspace MCP
- A key lets another app read one coven’s memory, facts, and assets. Per key, you can also let it suggest memories, which you approve in AICoven, and save drafts to Assets.
- Not offline
- Hosting an agent yourself changes where the agent runs. The model it uses may still be a cloud one, and its turns still pass through AICoven.
Setting up takes three steps.
- Download the appOn iPhone, iPad, or Mac. Sign in and Strix is there.
- Connect one AI accountOne is enough. On the Mac, Detect finds local agents and cloud credentials you already have.
- Ask for the first piece of workStart in Strix. When a client needs its own space, make a coven: a name, an icon, and the accounts and apps it may use.
Stuck on a step? The tutorial and the FAQ have the detail.
Give each world its own coven.
iPhone, iPad, and Mac. Your first coven is free.