Overview & result
agent-chaperone shows Agent Chaperone, built on Jev: Screen an agent's tool calls with Jev before they run.
From the original post
Agent Chaperone by @agentchaperone: Screen an agent's tool calls with Jev before they run.
How Jev fits in the loop
- An agent's tool call is screened before it runs, and the tool's result is screened before the agent reads it.
- Each screening question returns a probability.
- Thresholds in a policy file decide what happens; nothing is decided in a prompt.
- Every judgment is written to a local log; it starts in shadow mode, which blocks nothing.
How to reproduce
- Review the source documentation and repository at https://github.com/agent-chaperone/agent-chaperone.
- Configure environment with TypeSafe Jev / System One API keys.
- Run test prompts against the decision endpoints.
Why this build matters
Shows Jev in practical use for browser and autonomous agents.
Limitations
- Not a sandbox: it cannot stop a server from doing something the call did not describe.
- Not a guarantee: the README says adaptive attacks written against it will get through.
- Not local-only: screened arguments and results are sent to the model backend, with secret-shaped strings redacted first.