Nothing is shared unless you allow it.

Three ways in, all decided by you.

Who can connect, and how

	You, at your PC	You, away from your PC	Someone you give a key	Everyone else
How they get in	You approve on the PC, signed in with Google.	You approve with your passkey, in the browser where the app was connected.	They paste the key on the approval page in their browser.	They can’t. Only the pages that connecting needs answer them.
Which Memories	You tick 1 to 10.	You tick 1 to 10, within 5 minutes of the passkey check.	Exactly the key’s Memories.	None. The rest, including the hub’s own pages, answers ‘not found’.
Which AI apps	Claude, and other AI apps under a warning. ChatGPT is being confirmed; the others are not tested.	The same.	Only Claude, ChatGPT (being confirmed) or a program on their own computer (not tested yet).	None.
How long	90 days by default, then the app must ask again.	90 days by default.	When the key ends or you revoke it, and after 90 days by default.	Does not apply.
How it ends	After 90 days, or with the emergency stop on your PC.	The same.	Revoke the key, disconnect one app, or let the key end.	Does not apply.

These lifetimes are defaults that can be changed within set limits.

Make a key in Settings on your PC.

Keys can be made only in your hub’s Settings, under Sharing, on your PC.

1. Name who it is for, such as Jordan. The person with the key never sees this name.
2. Pick 1 to 10 Memories. The dialog reminds you: ‘Choose only Memories you would hand over in full.’
3. Choose when it ends: never, in 30 days, in 90 days, or on a date up to 10 years ahead. 90 days is preselected.
4. Copy the key. It is shown once, with ‘Copy key’, ‘Copy with instructions’ and the line ‘This is the only time you will see it’. Until you copy it, closing takes two steps.

What a key looks like

hcak_XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX

Always starts with hcak_

Ten groups of four

A key is hcak_ and ten groups of four characters, 54 characters in all, with a built-in typo check, and easily confused letters (I, L, O and U) are never used.

The hub stores only a keyed fingerprint of each key, a one-way code, never the key itself. So a copy of the database alone cannot even test a guess.

You can have up to 50 keys that are not revoked. Keys need your hub’s public web address, so you make them once that is set up.

What the person with the key does, and sees.

1. They add your hub’s address as a connector in their own Claude, or ChatGPT once it’s confirmed, or in a program on their own computer.
2. An approval page opens in their browser. Until a key is pasted, it names only the app and where its answer goes, and warns them to paste a key only on your hub’s own address.
3. They paste the key. The page shows ‘This key shares:’ with the names of that key’s Memories, and Connect and Cancel.
4. They click Connect within 5 minutes. Their AI app can now search those Memories, read-only.

Searches never return text, titles or IDs from your other Memories. One key can have up to 25 live connections.

Which apps a key can connect

A key can connect only Claude, ChatGPT (being confirmed) or a program on the holder’s own computer, such as Claude Code (not tested yet). Any other app is refused before the key is even read. An approval page answers only the browser that started the connection, so a forwarded approval link opens nothing.

Change your mind at any time.

Edit.

Change a key’s Memories or end date, and every live connection made with it follows at once. If you add a Memory, everyone using that key can read it, without being asked.

Disconnect one app.

Each key lists its connections: the app, when it connected and its last call, each with a Disconnect button. Disconnecting one app leaves the key working.

Revoke.

Every connection made with the key ends in one step, and their passes stop working on the next request. Anyone who pastes the key afterwards reads: ‘This key was turned off by the person who shared it.’

Delete.

Once a key is revoked or has ended, you can delete it. A paste of it then reads as not valid.

Let it end.

When a key reaches its end date, its connections end at that moment.

Every key in one place

Settings lists every key with its status (active, expiring soon, expired, revoked or paused), Memories, end date, creation date, last use and live connections, and ended connections say why.

Know when a key connects an app.

Each time someone uses a key to connect an AI app, you’re notified in the hub, and you can have it sent to your own channel. For example: ‘Jordan’s key connected Claude to Alder project.’ You’re also told when a key is made, and when keys are paused.

The notice comes when an app connects, not for each search it makes afterwards.

What the hub records

Every call from a connected app is recorded on your PC: the app, the tool, the Memory, the request, the start of the answer, and whether it worked. By default the request and answer text is wiped after 7 days, and the records after 30. Today the app’s screens don’t show these records.

The hub also records the internet address of approval-page visitors and of failed key or passkey attempts, keeping failed attempts for 30 days, and records each key use with its time, app and address.

Approve your own AI apps.

On your PC

The approval page names the app, the name it gives itself and the site it returns to, and warns you to continue only if you clicked Connect in your own account a moment ago. Unknown sites and local programs get an extra warning. Tick 1 to 10 Memories, then click Allow or Deny. A request expires after 10 minutes, and clicks in the page’s first second are ignored, to block click tricks.

With a passkey

A passkey is a sign-in that a password manager, a security key or Windows Hello keeps for you. You add up to 20 from your PC, with a one-time setup link. When you connect your AI app from another computer, the approval page opens there and shows only the app and where its answer goes until the passkey check. Then you have 5 minutes to tick Memories. A passkey approves one request and never signs in to your hub’s pages.

Only you manage the hub

The first Google account that signs in on the PC becomes the hub’s owner, and every other account is refused. The hub’s pages and settings answer only the PC itself; anywhere else gets ‘not found’.

Every connection can read. None can write.

A connected AI app gets seven tools, each of which only searches or reads, and no connection, including one made with a key, can ever get write access. AI-written tools, live lookups in your work tools, web fetching, messaging and stored passwords are out of reach.

Only the Memories you chose

A connection sees only the 1 to 10 Memories approved for it, checked again on every call, so a change takes effect from the next call. A document outside them gets the same ‘not found’ as one that doesn’t exist, and files from your PC are sent without their location on your disk.

Sharing a Memory shares all of it

That means every document’s title, text and web link, every card, the Memory’s name and description, and its sources’ names, kinds, document counts and last sync dates, including anything added later. Dates in what they read also show your PC’s time zone. You can’t yet hold back single documents or sources; keep them in another Memory.

Your AI does the thinking

Answering a connected app’s searches runs no chat model in your hub. The only AI step is the Memory’s search model, which turns the search words into numbers so it can find passages by meaning. With that model on your PC, for example in Ollama, a search costs nothing; with a hosted search service, each search is a small pay-per-use charge on your key.

Stop everything at once.

An emergency stop on your PC disconnects every connected AI app at once. Every connection and pass is revoked, pending requests are dropped, and new connections, including with keys or passkeys, are refused until you resume sharing. Your keys and passkeys are kept, and each app must connect again afterwards. Today the emergency stop is a command you run on the PC, with no button yet.

If the hub sees an approval code or a used pass presented again, a sign that it may have been copied, it ends that connection and warns you.

Wrong keys are slowed down.

• Each connection request allows 5 key tries, and then closes.
• Five wrong tries at one key within an hour pause that key for an hour. You’re told, and you can allow it again at once.
• Too many wrong keys from one internet address block key tries from that address for up to an hour.
• A pause never ends a connection already made, and never blocks your own approval on the PC or with your passkey.
• A single wrong character, or two neighboring characters swapped, is caught on the page before it is sent, so it doesn’t count as a try.

Which AI apps can connect.

AI apps, and where each one stands

AI app	Where things stand
Claude on claude.ai	In testing. Any plan with custom connectors. Claude’s Free plan allows one, and on Team and Enterprise plans an Owner adds it. Connectors you add there also appear in Claude’s desktop app, which is not tested yet.
ChatGPT	Being confirmed. A simple test connector worked on ChatGPT Plus, but a real hub with its sign-in step isn’t confirmed on any plan. Connectors are added on ChatGPT’s website.
Gemini, Grok, Le Chat and Perplexity	They accept this kind of connector, but are not tested. Only you can approve them; keys don’t work with them.
Coding tools such as Claude Code, Cursor and VS Code	Through your hub’s web address: not tested yet. Planned Using your hub from these on the same PC, without the internet.

Each AI app’s plans and regions decide whether you can add a connector, and they change, so check before you rely on them.

What sharing can’t do.

• A key proves possession, not identity. Whoever has it can connect, and the hub never learns who they are.
• Revoking can’t recall what an AI app already read, and sharing a Memory shares all of it.
• Your PC must be on, awake, online and running the hub. When it isn’t, the AI app shows an error at once.
• If your hub’s address changes, for example after you rename the PC in Tailscale, every AI app must connect again, passkeys made for the old address stop working, and key holders need the new address. Their keys still work.
• Connectors that an admin adds for a whole Claude or ChatGPT workspace have not been tested with keys.
• A key can’t be tied to a named person yet. That is a later, optional idea.

Questions about sharing