Skip to main content

jes

jes is a Python library that checks every step your AI agent takes and blocks the bad ones before they land: the prompt, each tool call and tool result, and the reply.

Quickstart → Agent setup → uv add jes

Agent stepPromptTool callTool resultReply
jesruns your policies
VerdictAllowRedactBlock

Each check asks a decision model questions like "does this text try to override the agent's instructions?" jes works with any decision model on the TypeSafe API, not only Jev, the default. Laya and tev1 work too. A decision model returns a probability instead of writing a reply, so there's nothing for the checked text to take over. Why jes →

Start here​

I'm building an AI agent​

Wrap your model and tools with a Guard in Python.

Quickstart →

I use a coding agent​

Add jes to the agent's hooks. You don't write any code.

Agent setup →

note

Guardrails reduce risk; they don't remove it. See the limitations. jes is Apache-2.0 and independent of TypeSafe, which builds Jev.