Jevex tries a different split for an AI agent. Jev decides what happens next at every step, whether to use a tool, ask a question or answer, and a language model is only called to write the content a tool needs and the final reply.
Python checks each tool call's arguments and Jev approves the exact call before three local tool servers run it, so the agent has twelve working tools and needs no outside accounts. Results go back into the state and the loop goes round again.
The repository includes a research benchmark that compares this loop with ordinary tool-calling loops on tokens and speed, and an illustrated explainer of how it works and where it falls short.
What you get
- Full Python source code under the MIT license
- Three local tool servers with 12 working tools
- The research benchmark and an illustrated architecture explainer
Jev
Python
FastAPI
JavaScript
Playwright
MCP
Hermes
