Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

tempo-jev-demo

Local natural-language task workspace where Jev, OpenAI or Gemini turn requests into view and task changes, with speed and cost comparisons.

AppprototypeMIT
View on GitHub ↗

Overview

Tempo is a local, natural-language task workspace for a fictional team. Users ask it to show work as a board, table, cards, bubbles, grouped canvas or calendar, then follow up to change the view, complete or reassign tasks, schedule events, capture notes or edit a checklist. A provider menu switches between TypeSafe's Jev, an OpenAI model and a Gemini model, each turning requests into workspace changes.

The author built Tempo with OpenAI's Codex and says they have not reviewed all of the code, calling it an experimental demo not to be used for anything serious. The README reports that in local tests Jev responded about 12 times faster than the OpenAI model and 14 times faster than Gemini, at lower cost, using published token prices. It ships with fixed fictional data and a fixed clock for reproducibility, plus eval runners with request and budget caps. The TypeSafe adapter can split large sets of independent questions into parallel batches.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich view or workspace action does this request call for?(inferred)board · table · cards · bubbles · canvas · calendar
ChoiceWhich task or person does the follow-up refer to?(inferred)—

How it works

  1. User types a request in natural language.
  2. The selected provider answers typed questions about the intended change.
  3. The app applies the change to the in-memory workspace.
  4. Eval runners compare providers on the same request sequences.

Strengths

  • Side-by-side provider comparison on identical requests.
  • Reproducible fixture data and fixed clock.
  • Budget and request caps on eval runs.

Caveats

  • Author has not reviewed all code; experimental only.
  • State lives in memory and resets on refresh.
  • Speed and cost figures are the author's local tests.

Getting started

npm ci && cp .env.example .env && npm run dev, then open http://localhost:5173.

At a glance

Type
App
Status
prototype
Language
TypeScript
Runtime
Vite + Express
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • I created Tempo with OpenAI's Codex, using Astra and GPT-5.6 Sol, with my direction and guidance. Codex also came up with the name Tempo. I haven't personally reviewed all of the code. This is an experimental demo so don't use it for anything serious. It may contain bugs, unexpected behaviour, and issues I'm not aware of. More details on how to get started below. Have fun!

    logicrw/awesome-jev-projects

  • A natural-language task workspace comparing performance across AI models (TypeSafe's Jev, GPT-5.6 Luna, and Gemini 3.8 Flash).

    hellogumbo/awesome-jev

Frequently asked questions

What does Jev decide in Tempo?

How a natural-language request maps to workspace changes, such as which view to show or which task to update.

How does it compare with the LLM providers?

The README reports Jev about 12 to 14 times faster and cheaper in the author's local tests.

Can I use it for real work?

No, the author describes it as an experimental demo with unreviewed code.

Related projects

  • json-renderer

    Playground comparing generative UI via an LLM writing json-render specs with Jev choosing and arranging prepared components.

    LibraryprototypeTypeScript
  • shapeshift

    One text box that morphs into the right UI card as you type, with Jev answering 14 typed questions per input and code doing the maths.

    DemoprototypeTypeScript
  • Cheshi

    macOS workspace for Codex with CodeGraph, Git and terminals, where opt-in Jev history recall finds past sessions and their decisions.

    Appbeta
  • jev-blindspot

    Prompt reviewer for Claude Code and Codex where Jev decides whether each prompt needs review before a light model lists blind spots.

    CLIbetaTypeScript
  • Qualm

    macOS menu bar screen-time app that asks Kev or Jev whether each screen breaks your rules and steps in with a pop-up.

    ServicebetaPython
  • zsh history completion with Jev

    X demo of zsh completion where Jev picks the most likely next command from the last 100 deduplicated history entries.

    Discussiondiscussion