AI Pulse · Under the Radar

Most AI assistants can draft an email or answer a question. The harder part is letting one do something useful in the real world, like book an appointment, order supplies, or handle a household task, without handing a language model your passwords.

Merit Systems has open-sourced OpenInstinct, a self-hosted iMessage assistant built around that problem. It can be deployed with one click and is designed to carry out tasks through messages. Its key design choice is an encrypted credential vault. Your login details sit in the vault, while the model is not meant to be able to read them. That is a different approach from closed assistant apps that ask users to trust a subscription service with access to sensitive accounts.

For business owners, this is worth watching because AI agents will only become genuinely useful when they can take action safely, not just produce text. The question is not whether an assistant can place an order. It is who holds the credentials, what permissions it has, and whether you can run the system on your own infrastructure. OpenInstinct is very early, with 202 GitHub stars and activity largely driven by its founder’s own post, so it is not proof that this model is ready for broad business use. But it points to a practical direction: assistants that act across tools while keeping sensitive access separate from the model itself. This is the kind of thing we build into an AI command centre, with clear controls around data, access, and actions.