The story
I run a trading system on a Windows VM. When something broke overnight, the routine was always the same: open remote desktop, find the right folder, read the log, check what changed. I wanted to ask my AI assistant instead, from wherever it runs.
So I built a bridge between the assistant and the machine. In one week of daily use it handled more than 4,400 tool calls. It also showed the real problem: 799 of those calls were unrestricted PowerShell and only 4 were fixed, pre-defined tasks, because nothing existed between read-only access and full control. Hostkeep is that middle ground, with walls the assistant can't climb.
Principles
- Least privilege first. Read-only by default, named folders only, tasks instead of a shell, no admin rights.
- Don't trust the AI. Assistants can be confused or tricked. Every protection is enforced in code on the machine.
- Evidence over claims. Every security claim names the code that enforces it and the automated test that proves it.
- Publish the gaps. What isn't finished is listed openly until it is.
- Your data stays yours. No Hostkeep cloud, no telemetry, logs on your machine.
How it's built
Hostkeep is built AI-first. Claude acts as architect and security reviewer: it writes the specifications and decisions and reviews every change. Coding agents implement against those written specs. Every fix starts as a failing test, and nothing is merged without evidence: the suite has 192 automated tests, plus stress tests that swap folders for links 100,000 times and a one-hour soak test.
Every design decision is recorded with the evidence behind it, including the ones that changed course after a test proved an earlier approach wrong.
Who's behind it
Fay, founder. Fay builds production systems by directing AI: a multi-account trading platform, signal processing, client websites and integrations. Hostkeep comes out of that work.
Hostkeep is not affiliated with Anthropic, OpenAI, Microsoft or any other company named on this site. Product names belong to their owners.
What's next
Human approval before changes or tasks run, separate keys per client, sign-in for web assistants, a one-click installer, and a pilot with three to five teams. The full roadmap is on the home page. If you run Windows workloads and want to be one of those teams, get in touch.