Technology
GreenPT is working on MCP with sign-in. What a connection with PrivaSearch would then offer: searching your client files with pseudonymised passages.
Not yet, but technically little stands in the way. PrivaSearch is an MCP server that an AI assistant signs in to with OAuth 2.1. As soon as the GreenPT chat can add an external MCP server with that kind of sign-in, it can search your own client files through PrivaSearch. According to its own plans, GreenPT is working on it.
GreenPT's roadmap lists under "Coming soon", for the API and the chat, improving and extending its MCP services with authentication (as of 29 October 2026). GreenPT gives no date, and does not say in which direction those services work: GreenPT offering tools to other assistants, or the GreenPT chat calling servers run by others. A connection with PrivaSearch needs the second.
How such a connection works, from the first request to the token, is set out step by step in What is an MCP server?. That it can work in principle is shown by the assistants that already do it: Claude, ChatGPT Business and Mistral Vibe connect with PrivaSearch in the same way.
GreenPT is a chat service that runs open models on its own servers. According to its own site, the chat is hosted in Europe and runs on renewable energy. There is no external model provider in between: what you enter stays with GreenPT. Next to the chat, GreenPT offers an API for open models.
That is a clear choice. Anyone who does not want to send conversations to a large American model provider will find in GreenPT an alternative that stays close to home. A comparison with other European services is in European alternatives to AI chatbots.
In the GreenPT chat you can give an assistant standing context with Spaces: guidelines, know-how and documents the assistant brings into every conversation. A document you add there or upload in a conversation goes in as it is. If it contains names, citizen service numbers, addresses or medical details, the model reads them too.
That is not a fault of GreenPT. It is how uploading works in every chat service: the whole file goes to the provider, including the pages you did not ask about. The difference with an American service is where the file ends up, not what is in it. What exactly happens when you upload is in Uploading documents to an AI chatbot.
With a connection, the client file stays where it is, and the assistant only receives the piece PrivaSearch returns. That piece is pseudonymised and checked by a colleague. The assistant can search, but not browse.
Opening a source is a second function, which returns one passage of at most 1,200 characters. No file, no file name and no link. If the search question contains a name after all, it is replaced before the search runs. The steps are written out on How PrivaSearch works.
PrivaSearch checks who is asking on every request. The token must belong to an active user with two-step verification, who is a member of the project and has completed the AI training. For a SharePoint library it checks live whether that user may still see the document there. If one check fails, the assistant gets nothing. Actions are logged, without document content in the log.
| Question | GreenPT with uploading | GreenPT with PrivaSearch, once the connection exists |
|---|---|---|
| What the model reads | The whole document, as it is | At most eight passages per search question |
| Personal data | Everything in the document | Replaced with pseudonyms |
| Who checks it | Nobody beforehand | A colleague releases every document |
| Rights per person | Whoever has the file can upload it | Per project, and with SharePoint live per document |
| Source reference | Depends on the model | Per passage, with a short reference |
| Trail afterwards | The chat history | A log of who asked for what and when |
| AI policy, DPIA and training | You arrange it yourself | Included |
GreenPT and PrivaSearch each solve a different part of the same problem. GreenPT decides where the model runs. PrivaSearch decides what the model gets to see.
With only a European model, the client file stays as it is: the model reads the names, just in a different place. With only pseudonymisation, the question remains where the model runs and what the provider does with the conversation. Together, a model on European servers only reads passages in which the personal data has already been replaced, from documents a colleague has released.
For the GDPR that matters. You do not arrange data minimisation with the location of the data centre, but with the question of how much goes to the model per request. The difference between pseudonymising and anonymising, and why pseudonymised data is still personal data, is in Pseudonymising or anonymising.
There is a practical advantage too. Anyone who already uses GreenPT does not have to switch to another assistant to search their own client files. The chat stays the same; only the source of the passages is added.
Two things stay with GreenPT, and ChatGPT, Claude, Mistral Vibe and the GDPR says the same for every connected assistant.
A third point concerns the model itself. An open model that a service hosts itself analyses differently from the heaviest closed models of the large labs. For one question that makes no difference, in a long analysis of a client file it sometimes does. A comparison on that point is on PrivaSearch or GreenPT?.
Go through these points before the first colleague switches a connection on.
The PrivaSearch connection works today with Claude (in the browser, in the desktop app and in Claude Code), ChatGPT Business and Mistral Vibe. Searching and opening sources also works in PrivaSearch itself, without an assistant. Which parties are involved in which link is in the sub-processor overview.
If GreenPT soon offers a chat that can add external MCP servers with sign-in, GreenPT is a candidate for the same connection. For a GreenPT user that means: the same European chat, with your organisation's client files behind it, without the model reading the personal data.
Want to see what an assistant gets to see through the connection? Request a demo.
No. For a connection, the GreenPT chat must be able to add an external MCP server and sign in with OAuth, and PrivaSearch must admit GreenPT as an assistant. GreenPT lists extending its MCP services with sign-in on its roadmap (as of 29 October 2026). Today the PrivaSearch connection works with Claude, ChatGPT Business and Mistral Vibe.
A chat that, as an MCP client, can add a remote server, with sign-in according to OAuth. That is how other assistants already connect with PrivaSearch: you sign in to PrivaSearch, approve the connection and the assistant gets a temporary token. In addition, PrivaSearch must admit GreenPT as an assistant; PrivaSearch decides that itself.
Not the documents themselves. GreenPT only receives the passages PrivaSearch returns, at most eight per search question, and those are pseudonymised and released by a colleague. Those passages do become part of the conversation and fall under GreenPT's retention policy. Your own question goes to GreenPT as well.
Because the location of the model does not change what it reads. A European model that receives a whole client file reads every name and number in it. Pseudonymising limits what goes to the model per request, and that is what the GDPR asks for under data minimisation. Pseudonymised data does remain personal data.
Yes, as soon as GreenPT can connect. Each connected assistant signs in separately on your organisation's connection address, and each connection can be revoked separately. Which assistants are allowed in your organisation is something you set down in your AI policy. PrivaSearch checks who is asking on every request, whichever assistant it is.
How retrieval-augmented generation works, how it differs from uploading and fine-tuning, why it reduces errors and which privacy questions to ask.
What the Model Context Protocol is, how a connection differs from uploading a file, what an assistant receives and which risks you should check.
What a language model does, what tokens and a context window are, why a model invents things and what it does and does not remember from your input.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →