Skip to content

MCP and privacy

What a connected
app can actually see.

When you connect an outside AI application to Kelven, it doesn't get your memory. It gets to ask narrow questions, and receives the minimum answer that makes the task work — usually with names and amounts replaced by typed placeholders.

Six stages, every request

Kelven minimizes the context it returns.

This pipeline applies only after a connected app asks Kelven for memory. It does not run on the message you typed into that app.

  1. Stage 1

    Authorization

    The request must carry a valid token with the scope for the capability it's calling. No token, no answer.

  2. Stage 2

    Relevance retrieval

    Only records related to the stated task are considered — never your whole profile.

  3. Stage 3

    Sensitivity filter

    Local-only records are removed absolutely. Records above your maximum sharing level, or switched off for connected apps, are removed too.

  4. Stage 4

    Context minimisation

    What survives is trimmed to the few facts the task actually needs.

  5. Stage 5

    Tokenisation

    Identifiers become typed placeholders — CLIENT_1, PROPERTY_1, AMOUNT_1 — with relationships preserved. Secrets are removed rather than replaced.

  6. Stage 6

    Result and receipt

    The client gets the minimised answer. You get a log entry showing what was returned and what was withheld.

How far a record may travel

You set the ceiling, per record.

  • PublicSafe to share anywhere.
  • NormalShareable with connected apps when relevant.
  • PrivateShared only when the task needs it, and tokenised.
  • ConfidentialNot shared with connected apps unless you turn it on per record.
  • Local onlyNever leaves the device. Never returned through the Gateway. No exceptions.

What we won’t claim

Kelven controls what Kelven shares.

If you type a client's real name straight into another AI application, that application already has it. No gateway can reach back and conceal it. What Kelvencontrols is its own half of the exchange: the context it returns, how much of it, and in what form.

The connected app may receive a minimised, tokenised version of the memory context it requested. That reduces what Kelven reveals; it does not reduce what you already typed directly into the app.

Every request appears in Connected Apps with what was returned and what was withheld, and disconnecting blocks the next request immediately.

Privacy Center · MCP Gateway overview

Your AI, your context. Start with Kelven today.

Put powerful AI to work with memory you control and protection before cloud requests.

Try Kelven