Skip to content

Chat With Feels

Most chatbots either sound flat or fake their feelings. Chat With Feels keeps a visible emotional state that changes how it talks - tone, cadence, emphasis - but not what it says is true. Not open to the public.

Web· tbd
Chat With Feels conversation beside a panel showing the assistant's current emotion levels
With feeling, without fiction. A chatbot whose mood is modeled in the open and changes its style, never its substance.

Overview

Chat With Feels is an experiment in a chatbot that has a mood and shows it to you. It keeps a running emotional state, displayed on screen, and that state changes how it talks to you - tone, cadence, emphasis and initiative. The facts and the safety rules stay the same whatever the mood.

Under the hood it tracks 22 emotion families, from joy and awe to confusion and nostalgia. Each one drifts back toward a calm baseline over time, and some pull on each other. A light reading of each message nudges the state, and the current state becomes style instructions for the language model. It also stores memories from past conversations and recalls the ones closest to what you're talking about now.

It is parked. It runs as a local build and isn't hosted anywhere public.

Why I built it

The chatbots I used were either flat - no sense of the room - or performative, faking feelings to keep me talking. I wanted a third option: an assistant whose emotional state is modeled, visible, and honest about being a model. The feelings get to shape how it talks. They never get a vote on the facts.

How it works

1. You write a message - a light pass reads it for emotional cues 2. The mood shifts - 22 emotion families move, then drift back toward calm over time 3. The mood sets the style - tone, cadence, emphasis and initiative, passed to the language model as instructions 4. The answer keeps its facts - the style changes; what is true and what is safe does not

Feature highlights

  • A visible mood - the assistant's current emotional state is shown on screen, not hidden.
  • 22 emotion families - from anger and fear to awe, gratitude and nostalgia, each fading back to baseline over time.
  • Feelings that interact - some emotions raise or damp others, the way anger pushes calm and joy down.
  • Style, not substance - the mood changes how it talks, never the facts or the safety rules.
  • Memory across sessions - it recalls the past moments closest to what you're saying now.

Notes

  • How it's built: A server-side emotion model feeding style instructions to a language model, with a web chat front end.
  • Tech stack: Node.js, TypeScript, OpenAI, MongoDB, a Vite front end.

You might also like

Illustration of a beaver damming a river in an overgrown ruined city, with a heron and ducks arriving at the new pond

Urban Beaver

Working prototype

Dam a river running through a ruined city, then watch it turn into a wetland. Gather logs and mud, stop the water, and see fish, frogs and herons move in. A different city each day.

Urban Beaver
Illustration of a person flying high above a real-looking city, river valley and mountains

DreamFlyer

Working prototype

Remember the flying dream, where you swim through the air? DreamFlyer puts it in a browser tab: stroke your way over Denver, Yosemite Valley or the Grand Canyon in Google's 3D map. No install, no account.

DreamFlyer
Agentic Directory showing a grid of AI tool listings, each with its score out of 36

Agentic Directory

Live

Need an AI tool that actually does the work? Describe your situation and get one honest pick, free. Agentic Directory scores more than 450 tools on a public 36-point rubric, and nobody can pay to move a score.

Agentic Directory