Jev in production › Agent tool and action selection

Home Assistant TypeSafe Conversation Agent

Home Assistant voice agent asks Jev typed yes/no questions per command to resolve intent instead of using an LLM, resolving commands in about 300ms (the-sof).

Open on GitHub ↗

300msmeasured, as published by the source
Use
Agent tool and action selection
Industry
General-purpose
Form
Open-source tool
Stage
In production
Plugs into
Home Assistant
Listed
2026-09-30
Found via
discord
Repository
the-sof/home-assistant-typesafe-conversation-agent
Stars
7
Forks
1
Last push
2026-10-11
Language
Python
License
MIT

The README opens with

A Home Assistant conversation agent that decides with a TypeSafe System One model instead of an LLM.

Status: early. Expect rough edges, and please report them. TypeSafe's hosted API needs a TypeSafe API key, which is metered (see Cost); a local server needs none.

A System One model returns typed, calibrated judgements rather than text. Jev is the one available today and the default; the integration is not written around it, so a later model is a config change. Every device command and state query is resolved from those judgements in ordinary Python, in a single API call. An LLM is used for exactly two things — splitting a request that contains several commands, and answering a general question — so the slow path is only taken when something...

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for agent tool and action selection