Explainability, right on the screen where AI speaks

When AI outputs need to be questioned, explained, or challenged, MUS puts explainability and feedback exactly where they happen. Embedded in any AI product with a face.

See MUS in action GitHub

Try it now

This is an interactive demo. Nothing you submit here is sent anywhere, no data leaves your browser.

your-ai-app.com / dashboard

Hover any card for 300 ms — then try voice feedback, thumbs, support, or video.

AI-generatedQuarterly Performance Summary

Revenue grew 17.4% quarter-over-quarter, driven primarily by expansion in the EU region. Customer churn dropped to 2.1%, the lowest in eight quarters. The model recommends focusing the next campaign on the mid-market segment, where engagement signals are strongest.

ForecastNext Quarter Prediction

Expected revenue€ 2.4M
Confidence interval87 %
Churn riskMedium

RankedTop Market Segments

Berlin Mitte94
Kreuzberg92
Prenzlauer Berg87

Why MUS exists

In AI products, interpretation is not binary. Two people can look at the same result and reasonably disagree. Context and feedback are not optional. They are the product.

We kept running into the same friction:

Users had questions about a specific section, not the whole app
Early testers disagreed with model outputs, and that disagreement lived in meetings instead of in the product
Feedback required extra steps: forms, tickets, context switching
For urgent matters, frustrated users were left to email or message developers directly

Traditional help systems are too detached. Traditional feedback flows are too heavy.

What if explanation and feedback lived exactly where interpretation happens?

That question became MUS.

What MUS does

MUS adds a contextual layer to any section of your web app. Wrap any output (a chart, an AI-generated summary, a recommendation, a score) with <FeedbackTarget>. Users hover and a toolbar appears, exactly where the output lives.

import { MusProvider, FeedbackTarget } from '@datachefhq/mus'
import '@datachefhq/mus/styles.css'

function App() {
  return (
    <MusProvider config={{
      projectName: 'My App',
      slack: {
        proxyUrl: '/api/slack-proxy',
        supportTeamEmails: ['you@company.com'],
        feedbackChannelId: 'C0XXXXXXXXX',
      },
    }}>
      <FeedbackTarget sectionId="ai-summary" sectionName="AI Summary">
        <AISummaryOutput />
      </FeedbackTarget>
    </MusProvider>
  )
}

Three things it does

Contextual video insight

Attach an explainer video directly to any section. Users get clarity right where confusion happens, with no redirects and no docs tab.

Frictionless voice feedback

Testers react immediately at the exact point of judgment, in their own voice. No forms. No references. No detours.

One-click support channel

A dedicated Slack channel opens instantly for urgent matters. Developers are reachable exactly where the issue surfaced.

What makes MUS different

Feedback happens inside the product, not around it. Every reaction is attached to the exact element the user was looking at. No screenshots. No “which page were you on?” The context is the message.

Voice carries what text loses. A 30-second voice clip captures hesitation, disagreement, and confusion in a way no form field can. Recorded in the user’s own voice, attached to the exact section.

Explanation that stays. Video insights aren’t a one-time onboarding tour. They live where the output lives, available six months in the moment a user asks “wait, why?”

An escape hatch for the urgent case. When something can’t wait, one click opens a dedicated Slack channel between the user and your team. No tickets. No bots. A real conversation, with full context already attached.

Your data, your destination. Feedback can route to Slack, Discord, Teams, or any webhook, with multiple destinations running in parallel. Self-hostable. Open source. No vendor lock-in.

Zero backend to write. Drop in the mus-server Docker sidecar and you’re done. No Node.js server to maintain, no auth integration to build.

What your team gets

Higher-quality tester signals: voice and in-context capture make feedback richer and easier to act on
Faster refinement cycles: disagreements with model outputs surface immediately, not in next quarter’s review
More engaged beta users: friction-free feedback raises response rates
Less ambiguity in model validation: every output has a clear feedback trail tied to the exact section
A direct line for urgent issues: frustrated users reach developers in seconds, not days

Built for AI products

MUS transforms interfaces from static displays into collaborative environments where outputs can be questioned, validated, and refined in context.

If your product output is interpreted, debated, or challenged by users, MUS is built for you.

If you’ve shipped a feedback form and watched it return one response per quarter, MUS is built for you.

If your team relies on quarterly user interviews to learn what testers think, MUS is built for you.

Common questions

Does MUS only work with React?

The component library is React (18+ and 19). The mus-server is framework-agnostic and works behind any frontend, so your backend doesn’t need to be React or even JavaScript.

Can I self-host everything?

Yes. The package is open source under MIT, and mus-server ships as a pre-built Docker image you run alongside your app. No SaaS dependency.

Does it have to go to Slack?

No. Slack, Discord, Teams, and generic webhooks are built in. Custom adapters take a few lines of code and can route feedback to Linear, Jira, your data warehouse, or any HTTP endpoint.

Does it work with non-AI products?

Yes. MUS is built around the idea of in-context feedback on specific sections, which is useful anywhere users interpret output. AI products are where it shines, but dashboards, reports, and internal tools benefit just as much.

What about user identity?

MUS auto-fills name and email from your auth system via pluggable resolvers (Stytch, Clerk, Auth0, NextAuth, or your own). If you don’t have auth, users can type their info or stay anonymous.

Is voice feedback really 60 seconds?

Yes, that’s the default cap and it’s configurable. The recording is converted to MP3 server-side and uploaded to your chosen destination automatically.

Philosophy

Contextual. Frictionless. Respectful. Iterative.

Small, precise moments of clarity, embedded exactly where interpretation happens.