Jev Chat

# Jev Chat

A chat-shaped command bar that calls real tools, **without an LLM writing anything.**



Every turn, a classifier picks: what was asked, which tool to call, which value goes in each
argument, whether to confirm first, and what kind of reply to give. Code does the rest: it calls
the MCP servers and builds the reply from the tools' own data. Because no model ever writes the
text, **the assistant cannot invent a fact**: every value on screen was either typed by the user or
returned by a tool.

An inspector pane shows the whole decision for any reply: the request, every question, every
probability, and what the code did with the answers.

- [Jev Chat](#jev-chat)
  - [What is Jev?](#what-is-jev)
  - [Run it](#run-it)
  - [How it works](#how-it-works)
  - [Adding a tool](#adding-a-tool)
  - [Layout](#layout)
  - [Notes](#notes)
  - [FAQ](#faq)
  - [References and resources](#references-and-resources)
  - [Contributing](#contributing)

## What is Jev?

[Jev](https://docs.typesafe.ai) is a classifier from TypeSafe. You hand it some state and a set of
named questions; it answers each one by choosing among options you supply. This app uses two of
its question types:

| Type       | Question                                  | Answer                                             |
| ---------- | ----------------------------------------- | -------------------------------------------------- |
| **Choice** | "Which of these tools fits?"              | one option key, with probabilities for all of them |
| **Noul**   | "Does the user say when the task is due?" | a probability that the answer is yes               |

Jev does not produce text. Every text argument, such as a search query, a task title or a city, is
copied from the user's message, the conversation or an earlier t