Connection
Assistant on MCP
- The question goes to the assistant the person pays for.
- We do not hold the login to that assistant.
- We connect only the sources you approve for this connection.
Security and data handling
The connection sends the question to the assistant the person pays for. The chat sends the question from the app on your address, and we hold the login to it. In both cases a public model writes the answer.
Data path
For both services we keep a search index of excerpts. We hold the team login only for the chat. With the connection, people sign in to the assistant they already have. That does not replace the originals.
How Anthropic or OpenAI handle data is set by their terms. We do not take those terms on for them. A wrong figure in an answer is visible from the source citation, and a person checks it. A public model is what keeps the listed price. A private model of comparable quality would raise it sharply.
The difference
Connection
Chat
01The question and excerpts leave
The model receives the question and the excerpts that fit it. Not the whole archive.
02Originals stay with you
Pohoda, the database, email, and files stay with you. Contracts and exports stay in Drive and OneDrive too.
03Both services have an index
From the approved sources we build a search index of excerpts, a vector database. A new database row, a new email, and a new file show up after the index refreshes. The interval is agreed per source.
04We don't take on the model's terms
What the assistant's provider does with the excerpts is set by their documentation. We are not liable for the provider.
05A wrong figure shows in the citation
The answer points at the source. If the figure does not match, a person checks it. The model is not the authority for it.
Trust
We take one source you want connected first and go through what leaves it for the model, and what stays with you.
Write a message