Connecting your sources is the easy part. Deciding what to connect isn't.

SharePoint, Slack, Confluence, Google Drive and Notion now connect to RAGSuite directly — in the open Community Edition, under Apache 2.0. Point one at a source, authenticate, choose what to index. It takes minutes.

That sentence is the whole product announcement. The more useful half of this post is what it changes about the conversation — because when the wiring stops being the hard part, the hard part turns out to have been somewhere else all along.

The gap that made self-hosted AI theoretical

For the last year the sovereignty argument has had a practical hole in it.

You can agree entirely that your company’s knowledge should not be processed on someone else’s infrastructure. You can have the servers, the budget and the mandate. And then you look at where that knowledge actually is: a decade of process documents in SharePoint, the decisions that never made it into a document sitting in Slack threads, specifications in Confluence, and a shared drive that everyone is slightly afraid of.

An AI platform that cannot read those is an AI platform for a pilot project. It answers questions about the documents somebody had time to upload, which is not the same thing as answering questions about how your organisation works.

So the practical objection to self-hosting was rarely about sovereignty. It was: the convenient product already reads our SharePoint, and yours doesn’t.

That objection is now closed.

What actually shipped

Five in-house connectors, alongside the ones that were already there:

• SharePoint — the document estate most DACH enterprises actually run on
• Slack — where the reasoning behind decisions tends to live
• Confluence — specifications, runbooks, internal documentation
• Google Drive — the shared-drive layer
• Notion — increasingly the wiki of choice for younger teams

Plus what was already in place: Gmail, website crawling with job monitoring, direct document upload, an open Model Context Protocol server and client — that is the emerging open standard for attaching AI systems to tools — and the n8n automation integration.

All of it is Community Edition. Apache 2.0, unlimited users, no connector reserved for the paid edition. We have been explicit that connectors would never be a paywall, and this is the release where that stops being a promise and becomes a property of shipped code you can read.

The same release refreshed the product interface to match the RAGSuite design system — quieter, more consistent, and considerably less like three products that grew next to each other.

The part nobody puts in a launch post

Here is the thing we would rather say plainly than have you discover in week three.

A connector reads what the account it authenticates with is allowed to read. That is the correct behaviour — an AI system that could bypass your permission model would be a security incident, not a feature. But it has a consequence that catches people out.

Most organisations’ access permissions are not a designed system. They are sediment. A site opened up for a project in 2021 and never closed. A channel that was private until someone needed to add a contractor. A shared drive where the folder structure encodes an org chart from two reorganisations ago. None of that caused visible harm, because finding something you were not meant to see required knowing it existed and going looking.

A search box that answers questions in natural language removes both of those requirements.

This is not an argument against connecting your systems. It is an argument for connecting them deliberately — and for treating the first index as what it genuinely is: the most thorough permissions audit your organisation has ever run. Most teams find something. Better to find it this way.

None of those are technical questions. That is the point: the technical part is now genuinely short, so the governance part is the project.

Why this is a sovereignty story, not a features story

It would be easy to file this under integrations and move on. But the reason it matters is narrower than “more connectors is better”.

The trade every convenient AI product offers is: give us access to your knowledge and we will make it searchable. The access is the price. Once your SharePoint is being read by a service you do not run, the location of your data has been decided by a contract rather than by architecture, and the exit question — what happens to all this if the vendor changes terms, gets acquired, or simply stops — is no longer yours to answer.

What changed here is that you can now take the same convenience without the trade. The connector runs inside your deployment, reading your systems, producing answers with citations back to the source document, on hardware you control. And because the code is public, “nothing leaves your network” is something you can check rather than something you have to be told.

Where to start

If you already run RAGSuite, the connectors are in the Community Edition — update and they are there.

If you do not, the Community Edition is public on GitHub under Apache 2.0 and installs from npm. There is no account, no trial key and no form in front of it; the connectors page lists what attaches today, and the changelog records what shipped and when.

Start with one source. Preferably a boring one. The interesting part will not be the technology.

Frequently asked questions

← All posts