Knighta guard that sits between an AI agent and the world

Check what an agent is about to do

Pick an example or write your own action. Knight reads what it really does, not what it looks like, and decides whether to allow it, hold it for a human, or block it.

Knight remembers what a session has done, so a risky pair of steps is caught even when each step looks fine.

Nothing checked yet. Try “Read the password file” on the left, then the same request hidden in base64.

Recent decisions

    Review queue

    Anything Knight held or blocked waits here. Approve it and Knight learns it was fine; keep it blocked and Knight learns it was not.

    Does adapting help?

    A fixed set of hand-written cases, the same on every run. Each row shows how many got through three versions of Knight: one that never learns, one before it learns, and one after it learns from the misses in the “allowed programs misused” row only.

    Charter

    The charter is the agent’s authorization boundary: what it may touch, where it may connect, how much it may spend. Edit it and Knight enforces the new version immediately.

    Audit log

    Every decision, review and adaptation is chained to the one before it. Changing or deleting any past entry breaks the chain.

    Resets the learned model, rules and audit log in this page.
    Knight
    Loading the engine (about 10 MB, first visit only)
    Knight runs entirely in your browser. Nothing you type leaves your device.