A Windows app for using AI with your own knowledge

It’s your knowledge. You decide who can use it.

Connector Hub gathers what you know from files, folders, web pages, and work tools such as Confluence, Jira, and Notion into searchable collections called Memories, stored locally on your PC.

You control which Memories your AI apps, such as Claude, ChatGPT, or any other AI application that supports MCP (Model Context Protocol), can access. You also decide who receives an access key, which grants read-only access to only the Memories you select, until the key expires or you revoke it.

Connector Hub is a desktop app you use in your browser. It runs on your own PC, and so does the MCP server your AI apps connect to (in testing). Everything it keeps stays there. There is no Connector Hub cloud: the app sends nothing to us, and you can limit what leaves your PC.

See how it works Try the demo

Nothing to download yet. The core app is built, and sharing with AI apps and people is in testing.

Example only, not a real key

Access key

For: Jordan
Opens: Alder project
Allows: Search and read
Ends: In 90 days

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

Illustration. A real key is shown only once.

Seven questions, in the order people ask them

The 30-second version.

What is it?
A Windows app that keeps your knowledge in searchable Memories on your own PC.
What does it do for me?
AI answers from your own documents. In the hub’s own chat, each answer lists its files.
How does it work?
Your PC searches your Memories. An AI writes the answer: the model the hub uses, or your own AI app (in testing).
Where does my data go?
Your Memories stay on your PC. Some content leaves it, such as what your AI app asks for. Nothing is sent to us.
Who can use it?
Nothing is shared unless you allow it: you approve your own AI app on your PC or with your passkey, or you give someone an access key for Memories you chose (both in testing).
What do I need?
A Windows 11 PC (64-bit) that stays on, an AI app that accepts custom connectors, such as Claude, a Tailscale account and a Google account. ChatGPT is being confirmed.
What can’t it do yet?
You can’t download it yet, and sharing with AI apps and people is still in testing.

How it fits together In testing

Knowledge on your PC. Thinking in your AI. Keys in your hand.

Your Memories stay on your PC. Your AI app asks, the hub searches only what you allowed, and your AI writes the answer.

An app you approved
A key you handed out
A key you turned off
Not shared

Your knowledge Built

Bring what you know. Get answers with their sources.

Collect documents into Memories on your PC and ask. The hub’s own chat lists the files behind each answer.

Three words to know

Memory.: One collection of knowledge, for a topic, team or client, with its own search.
Source.: Where a Memory’s knowledge comes from:

Files you upload

A folder on your PC

Web pages

Confluence Cloud

Jira

Notion

GitHub
Hub.: The app on your PC that keeps and searches your Memories.

Which sources the first release will include isn’t decided yet.

Two ways to ask

In your own AI app In testing

Claude searches the Memories you allowed and writes the answer. ChatGPT is being confirmed.

In the hub’s own chat Built

A model on your PC through Ollama, or an online one with your own API key, writes the answer.

Connector Hub · Chat

What is our refund window for annual plans?

Annual plans can be refunded within 30 days. After that, a team lead can approve an exception.

Files used

Refund policy.pdf, Policies folder
Exceptions, Confluence Cloud

Illustration of an answer in the hub’s own chat. Small AI models may cite less.

Your data

Stored on your PC. Here is what leaves it.

These are the main times something leaves your PC. Nothing is sent to us.

Stays on your PC

Copies of your documents and uploads
The search index and every version of each document
Chats, cards, settings, records and your passwords for services

Leaves your PC, and when

To an AI app you or a key holder connected In testing: the results and document text it asks for, from the Memories allowed.
To the hub’s own AI providers, if they aren’t on your PC: your question, matching passages and recent messages; for a correction, also the answer, your comment and the text around its cited passages; and, about once a day unless you turn tool ideas off, a summary of each Memory with document titles and recent questions. When only models on your PC are switched on, none of this leaves it.
To an online search service, if you pick one: every passage and search.
To your messaging apps: notices and results. Most notices go to every app you connect by default; a disagreement notice includes your question and comment.
To your work tools: requests made straight from your PC, with the credential you gave.
To an OpenAI-run address, when the hub first indexes: a request to download a public word list it uses to measure passage sizes, with no document content.

Sent to us

Nothing.

There is no Connector Hub cloud. The app sends nothing to us, and we run no server that holds your data.

Where your data goes, including Tailscale and Google sign-in

Access keys In testing

Give access you can take back.

For example, a consultant gives a client a key to that client’s Memory, set to end when the project does.

Only the Memories you pick.: 1 to 10 of them, never your other Memories.
Read-only, always.: Nothing can be added, changed or deleted.
Ends when you say.: Never, in 30 or 90 days, or on a date you pick. Revoke it sooner at any time.
You’re told when it connects.: 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.

Good to know

A key works for whoever holds it, and the hub never learns who they are. Give each person their own key, and send it privately, as you would a password.
Revoking ends access, not copies. What an AI app already read stays with that person’s AI provider.
Sharing a Memory shares all of it, including documents added later. Keep anything private in a Memory you don’t share.

How sharing works, step by step

Try it Demonstration

Make a demo key.

Make a key, let someone connect, change what it opens, then revoke it. Nothing here is real, and nothing is sent anywhere.

No real key is made here. Real keys are made only on your PC.

One key, two sides
What you do	What you see	What they see
Make a key for a client that opens Support handbook and Alder project, ending in 90 days	The key, once: ‘This is the only time you will see it.’	Nothing yet
Send it privately	Nothing changes until they connect	The key, in a private message
Wait while they paste it and click Connect	A notice: ‘A client’s key connected Claude to Support handbook and Alder project.’	‘This key shares:’ Support handbook, Alder project
Untick Support handbook	The key now opens only Alder project	From its next request, their Claude can’t search Support handbook
Revoke the key	Revoked, 0 live connections	Their Claude loses access. If they paste the key again, they read: ‘This key was turned off by the person who shared it.’

What you didn’t have to do

Move your documents to another online service. They stayed on your PC, and each AI got only what it asked for, from the Memories you picked.
Hand over your Confluence or Jira login. Connections can’t reach your stored credentials.
Show them everything. The key opened only the Memories you picked.
Trust the key forever. You could give it an end date, and revoke it at any time.

Your own AI app In testing

Connect your own AI app. You approve it.

Add your hub’s address to your AI app as a connector, then approve it and tick the Memories it may read.

At your PC.: Approve in your PC’s browser, signed in with Google.
Away from your PC.: Approve with your passkey, in another computer’s browser.
Someone you gave a key.: They paste it on the approval page and get exactly that key’s Memories.
Everyone else.: Anyone else at your hub’s address reaches only the few pages that AI apps and approvals need. The rest answers ‘not found’.

For every connection

An approval request you don’t answer expires after 10 minutes.
Connections end after 90 days by default, and the app has to ask again.
An emergency stop on your PC disconnects every AI app at once.

More about approving

What you’ll need

A Windows 11 PC that stays on, and an AI app that accepts custom connectors, such as Claude.

The people you share with need only their own Claude account, or ChatGPT once it’s confirmed. They install nothing.

A Windows 11 PC (64-bit).
Windows 10 and ARM PCs are not tested. A Mac version is planned.
On, awake, online and running the hub.
Your AI can reach your knowledge only then, and after a restart only once someone signs in to Windows and the hub is running again. So leave the PC running during working hours.
An AI app with custom connectors.
Claude on claude.ai is in testing; Claude’s Free plan allows one custom connector. ChatGPT is being confirmed.
Which AI apps can connect
A Tailscale account. In testing
It gives your hub its web address. Sign in to Tailscale with a personal login, such as Gmail or Apple, not a work account. Tailscale says its free Personal plan is for non-commercial use, so businesses should check its paid plans.
A Google account. In testing
You sign in to your hub with it, on the PC, so only you can manage the hub. Google is contacted only while you sign in.
Optional: Ollama.
It runs AI models on your PC, so search by meaning and the hub’s own AI work can stay there.

How setup will work Planned

One signed installer, one administrator prompt and two clicks in your own Tailscale console. None of this exists yet.

Where things stand

What’s built, in testing, planned and not decided.

Nothing is released yet. Built means it works in our own copy of the app today. In testing means it is built and tested in a separate version, not yet in our own copy. Planned means designed, not built.

Built

Memories from files, folders, web pages and work tools
Search by words, and by meaning with a search model
Cited answers in the hub’s own chat
Corrections: fix a wrong answer once
Scheduled questions and messaging outputs
AI-written tools

In testing

Google sign-in for the hub’s owner
Connecting your own AI app
Approving with a passkey
Access keys

Planned

A signed installer and setup wizard
A tray icon, and starting with the PC
Automatic Tailscale setup
A keep-awake switch for sharing
Same-PC use without the internet
An optional update check, off unless you turn it on
A Mac version
Later and optional: keys tied to named people, and holding back documents in a shared Memory

Not decided

Price
Release date

Limits today

Stopping all sharing at once is a command on the PC, with no button yet.
It can’t read scanned PDFs, images, audio, video or old Office files such as .doc.

By design

There is no cloud copy to fall back on. When the PC is off, your AI can’t reach your knowledge.
Your hub’s web address is public and appears in public certificate logs, so give the PC a neutral name.

Questions

Short answers to what people ask first.

How does Claude reach my knowledge if it stays on my PC?

In testing: through a public web address from your own Tailscale account. What Claude reads goes to Claude’s servers, under your account. Tailscale says its relay cannot decrypt your hub’s traffic. It does see connection details: your hub’s address, the addresses that connect, and how much data flows and when.

Is my knowledge encrypted?

Not by the hub. It encrypts only your passwords and keys for other services, and the encryption key sits in the hub’s folder next to the database. That protects a stray copy of the database, not against anyone or any program that can use your Windows account, and a backup of the folder holds the key too. Use device encryption such as BitLocker.

Will my documents be used to train AI?

Not by us: the app sends us nothing. What an AI provider keeps or learns from is set by the account it is used under, yours or a key holder’s, and that account’s settings.

Can I keep all AI work on my PC?

Inside the hub, yes: with Ollama and its models on the same PC, every model job can run locally. Put the local model first on the AI models page and switch off any online provider, or the hub can fall back to it. A connected AI app is different: what it asks for goes to its provider.

Do I need an API key?

Not to connect your own AI app (in testing): the hub runs no chat model for it. Only AI work inside the hub needs an API key, or a model on your PC.

Does it work with ChatGPT Plus or Free?

Not confirmed yet. A simple test connector worked on ChatGPT Plus, but a real hub with its sign-in step isn’t confirmed on any plan. OpenAI’s help pages don’t list Free for custom connectors.

What will it cost?

We haven’t set a price yet. Your AI plan, an API key or a paid Tailscale plan for business use can cost money. The hub estimates AI costs but doesn’t cap them.

Can my business use it?

The draft license allows use for your own work or business, on computers you own or control (an employer’s only with its permission). It is a draft for legal review, not in effect. Tailscale says its free Personal plan is for non-commercial use, so businesses should check its paid plans.

Is it open source?

No. It will be closed-source, under a license that is still a draft. Its open-source parts keep their own licenses.

Words we use

Plain meanings for the terms on this site.

Memory: A collection of knowledge with its own sources and search.
Source: A place a Memory learns from, such as a folder or a work tool.
Card: A short fact in a Memory that ranks higher in search.
Correction: The hub’s fix to its own copy after you press Disagree and say what’s right.
Access key: A code that opens chosen Memories, read-only, until it ends or you revoke it.
Passkey: A sign-in kept by a password manager, a security key or Windows Hello.
Connector: An outside service you add to an AI app, such as your hub.
MCP: Model Context Protocol, a standard way for AI apps to use outside tools.
Tailscale and Funnel: Another company’s network service, and its feature that gives your PC a public web address.
Search index and search model: An index finds passages by words; a search model lets the hub find them by meaning.
API key: A secret from an AI provider that lets the hub use its models, billed per use.
Ollama: A separate program that runs AI models on your own PC.
Token: A password-like code from a work tool, or a piece of text that AI providers count and bill.