Claude has shipped the first MCP integrations.

· Bits and Bobs 5/5/25
  • Claude has shipped the first MCP integrations.
    • Unsurprisingly they're going with more of the app store model.
    • There's a small set of approved MCP integrations you can enable.
    • The integrations are all aimed primarily at enterprise cases.
    • They've also only allowed the integrations in the Max subscription.
      • When you're worried about the downside risk of a feature and want to experiment to see how bad it is in the wild, a classic technique is to roll it out to a very small audience and watch carefully.
      • Presumably the number of users with a Max subscription is many orders of magnitude lower than their total user count.
    • The enterprise focus also tends to focus on things that have less prompt injection risk.
      • Things that are pulling from data from inside an enterprise are more likely to pull from data that was written by employees and thus more trustworthy than, say, emails.
      • But there are many internal systems that allow untrusted content.
      • For example, it's not uncommon for user feedback flows to automatically create JIRA tickets.
    • The main danger of MCP is not misbehaved integrations (though that is also a worry), it's prompt injection.
    • MCP is great for things with data entirely inside the house (only your employees, not injectable) and/or things that can't have irreversible side effects.
      • But lots of things have untrusted data (e.g. auto-filed JIRA tickets) or surprisingly have irreversible side effects (e.g. any network request).
    • Prompt injection can happen even for a well-behaved integration, for sources that allow open-ended or untrusted inputs (like search results, emails, etc).
    • Limiting to a subset of trusted MCP integrations does not meaningfully mitigate prompt injection.
    • The app store model leads to gatekeepers, but doesn't address prompt injection.
    • So now there are two problems!

More on this topic

From other episodes