Buying guide
How to choose an MCP server for company knowledge
Connecting an MCP server takes minutes. Deciding which one speaks for your company deserves longer, because whatever it returns, your AI tools will repeat with confidence.
In short
Check five things before connecting an MCP server for company knowledge: whether it enforces each user’s permissions, whether every answer carries its source, how it stays current and dates its facts, where the data is hosted and processed, and how reads are limited or priced. A server that fails on permissions or sources will leak or mislead.
Updated
Permissions come first
An MCP server answers on behalf of the person whose AI is asking. It should return only what that person may see, and a revoked share should stop returning data on the next request, not at the next sync.
- Is access scoped per user and per workspace, or is it one shared key?
- What happens to an agent’s access when a share is revoked?
- Can the owner see what was read through a share?
Every answer needs a source
A server that returns fluent summaries without sources asks you to take it on trust. Look for records that name their source and link back to it, so a doubtful answer can be checked in the meeting or thread it came from, and for an explicit status on each record, so a proposal is never presented as settled.
Ask how it stays true
Most knowledge bases go stale because someone has to update them. Ask how new information gets in and how quickly, and whether each record carries the date it was true as well as the date it was captured. Ask too how a wrong record is corrected, and whether the correction applies for everyone or only in one chat.
Hosting and read limits
For European companies, check where data is stored, where model inference runs, who the sub-processors are, and whether your data trains anyone’s model. Then check the price of reading: AI tools and agents read far more often than people do, so per-read charges or low caps can surprise you.
- In Parallel: hosting and model inference in Frankfurt, on Amazon Web Services and Amazon Bedrock, with every sub-processor in the EU
- In Parallel never trains models on your data
- Reads are free up to 10,000 per user a month, pooled; above that, each read uses 0.1 Knowledge Units
FAQ
Common questions
- What is an MCP server?
- A service that exposes data or tools to AI assistants through the Model Context Protocol, an open standard. Once it is connected, MCP-capable clients such as Claude, ChatGPT, Copilot and Cursor can read from it.
- How should an MCP server handle permissions?
- It should answer as the user whose AI is asking and return only what that user may see. In Parallel scopes reads to workspaces, and when a share is revoked the next request returns nothing.
- Do MCP reads cost extra with In Parallel?
- Not within fair use. Reads from your AI tools and agents are free up to 10,000 per user a month, pooled across the company. Above that, each read uses 0.1 Knowledge Units.
Start with your next meeting.
Bring your team in. Use every surface, including every other team's. From €29 per user per month, billed monthly, no commitment.
30-minute demo. We'll show you how meeting signals become a shared record of decisions, owners and risks.