Meet Jenny: An AI Assistant You Can Actually Trust

Personal AI assistants are having a moment. You message one from your phone, and it checks your calendar, reads your email, searches the web and gets on with the job. The idea is great. The worry is obvious too: something that can do anything for you can also do the wrong thing, or be tricked into it. We’re building our own, and we’re calling her Jenny.

She isn’t finished. We’d rather say so up front. This post is about what Jenny already does well, why we built her, and what’s still ahead.

The idea

Jenny is a personal assistant that lives on a Raspberry Pi, a small, cheap computer that sits on a shelf and stays on. You talk to her on Telegram or from the command line. She remembers what matters, reminds you of things, looks things up, and works on tasks you hand her.

The idea isn’t new. Open projects like OpenClaw showed how useful an always-on assistant can be, with plain-text memory files, a regular “check in” routine and plenty of chat channels. We learned a lot from them. Then we asked what we’d change, and the answer came down to one word: control.

Where we think we’re improving the idea

She asks first, and the rules are enforced by code. Jenny starts with no permissions at all. When she wants to read your calendar or edit a file, she asks. You can allow it once, say no, or choose “save this permission”, and she turns your answer into a clear rule in plain English. The important part is who enforces it: the program, not the AI. The AI can suggest a rule but can never grant itself one. Even if a sneaky web page or email tells her to do something she shouldn’t, nothing happens unless a rule you approved allows it.

Your private data doesn’t leak out through the back door. A common trick against assistants is hiding instructions in a web page that make them send your information to a stranger’s website. Jenny has a guard for exactly this. Once she’s handled something private, like mail or your files, she won’t fetch from a new website unless you’ve approved that specific site.

Memory you can read, and that you approve. Jenny keeps a searchable log of every conversation, plus a short set of notes about you in ordinary text files you can open and edit. New facts go into the notes only after you say yes. When something relevant from the past comes up, she brings it in on her own.

Not tied to one AI company. Jenny works with Claude, OpenAI, Kilo Gateway and Ollama. You pick which model does which job, and set a backup order. If your first choice is down, she moves to the next and tells you. She can even use a small model on your own network to sort easy questions from hard ones, so you only pay for the big model when it’s needed.

A spending limit that really stops. Every AI call is counted and priced. Set a daily budget, get a warning at 80%, and at 100% Jenny stops calling models until midnight. No surprise bills.

Safe when she works on her own. Shell commands run inside a sandbox. Reminders and scheduled jobs only happen because you asked. A regular “heartbeat” check-in, where she works through a to-do list you wrote, is off until you switch it on. Even then, she only does what her permissions allow.

She gets better at your routines. When Jenny finishes a long task, or you correct her, she can write down how it was done as a short step-by-step skill. Skills can’t run code or add new powers. Every step still goes through the same permission checks. The worst a bad skill can do is give bad instructions to a gate that still checks every move.

What she can do today

  • Chat on Telegram and in the terminal, sharing one memory and one conversation.
  • Read and write Google Calendar and Gmail, with your permission.
  • Search the web and fetch pages.
  • Read, edit and organise files in folders you choose.
  • Run commands in a sandbox.
  • Set reminders and recurring tasks.
  • Work on your coding projects: read the code, make edits, run the tests, and commit. Pushing code always needs your explicit go-ahead.

Who it’s for

People who want an assistant that does real work but who don’t want to hand over the keys to everything. If you’ve thought “this would be useful if I trusted it,” Jenny is being built for you.

Where we are, honestly

Most of the foundations are in place and working: permissions, memory, scheduling, Google, the sandbox, skills and coding mode. Still ahead are polish and hardening, a simple web console for your home network, and more real-world use. Jenny speaks Telegram and the command line today. She doesn’t have the dozens of chat channels some other projects offer, because we’d rather get the trust part right first. Some details above may change as we learn. We won’t put a launch date on her until we trust her.

If you’d like to follow along, or have a job you’d want an assistant like this to handle, get in touch. Your use case may well shape what we build next.

Wondering whether AI is right for your business? Read How Can AI Actually Help Your Business?