Setup required Set your Worker URL in config.js (or here) to start talking. The NVIDIA key stays in Cloudflare — it is never part of this site.

Unleash the power of

ThinkingMachines

Q: How do I start a conversation?

Press Start talking and allow microphone access, or type a message in the composer. Replying aloud can be turned off in Settings.

A real-time voice interface for NVIDIA-hosted models. Speak naturally, hear the answer out loud, interrupt any time.

Space mic Esc end mic → ASR → LLM → TTS

Idle — start a conversation

Microphone is off.

Connection error

Mic opens only during a session.

model — asr — tts — Scroll

Configuration

Settings

Audio devices, privacy and runtime — saved to this browser. The agent’s name and instructions are fixed.

01 Connection

This site is static (GitHub Pages), so requests go through your own proxy. The API key lives in the proxy, never in this page.

Proxy base URL Pre-filled from config.js at deploy time. Override here only for local testing; the value is stored in this browser. Leave blank on localhost to use the bundled dev server at /api.

App token (optional) Lets your proxy refuse strangers after you publish the page. Stored in session storage only.

Model ID

Only models your NVIDIA account is entitled to will run — the catalog lists far more than it allows. openai/gpt-oss-20b is verified working here (~3s per turn); if a model 404s the app retries once with it automatically.

untested not configured

Temporary one-off key? Use the local dev server (npm start) or your Worker's secret — do not paste keys into this page on a public host.

02 Agent

Agent name

System instructions Fixed by this deployment and sent as the system message on every turn. Spoken replies stay short by default.

Response style

Language

locked

03 Audio

Microphone

Speech output voice

Microphone test

idle

Bars move with live input level. Permission is requested when you test or start a conversation.

Speech rate 1.0×

Pitch 1.0

04 Privacy & storage

Show transcriptDisplay the live transcript beside the orb. Save conversations on this deviceTranscripts only — audio is never recorded or stored. Turn off to keep sessions ephemeral. Speak responses automaticallyUses your browser's speech synthesis engine. Barge-inCut the agent short the moment you start talking again.

Audio: Processed in the browser; not saved.
Transcripts: localStorage on this device, if enabled.
Transport: HTTPS to your own proxy.

05 Runtime

Host: —
Proxy: —
ASR: —
TTS: —
Model: —
Storage: —

Browser speech recognition handles transcription; the NVIDIA model handles reasoning and wording; your browser speaks the reply.

About

About Remy Agent

A real-time AI voice and chat assistant by DotcomAI, powered by NVIDIA-hosted inference through a proxy you control. Here is how it works and what it stores.

01 Pipeline

Microphone → browser speech recognition (ASR) → your proxy → NVIDIA chat-completions model → spoken reply through the browser's speech synthesis engine (TTS).

Turn detection comes from the recognizer's own segmentation plus an amplitude gate on the mic stream. Interrupting the agent cancels playback and the in-flight request immediately.

Model: Yours to choose — any chat-completions endpoint on your NVIDIA account.
Audio: Never sent to the LLM, never written to disk.
Limitation: Native speech-to-speech is not used: speech recognition and synthesis are browser-native, so the model sees text.

02 What is stored

Settings live in localStorage. Transcripts are stored locally only when you leave "Save conversations" on, and you can delete them individually or all at once.

Your NVIDIA API key is never part of this page, its bundle, or this repository. It sits in your proxy's environment. If you use the optional app token, it is held in sessionStorage and disappears when the tab closes.

Your proxy and NVIDIA process the conversation text under their own retention policies — this app cannot make claims about those.

03 Accessibility

Every control is a real button with an accessible name and visible focus. Conversation state is announced through an aria-live status region and mirrored in the header chip. Reduced-motion preferences disable orb animation. Touch targets stay above 44px on mobile.

04 Deploying

Static files only — no build step. Push index.html, app.js and .nojekyll to a repo, enable GitHub Pages on the branch root, then paste your proxy URL in Settings.

The same repo ships worker.js (Cloudflare Worker) and a zero-dependency server.mjs for local testing.

05 Frequently asked questions

What is Remy Agent?

Remy Agent is a real-time AI voice assistant built by DotcomAI. You speak or type, and an AI model answers — spoken aloud and shown as a live transcript.

Is Remy Agent free to use?

The web app itself is free to use. Model access depends on the inference account behind this deployment, including its rate limits and quotas.

Can I choose different AI models?

Yes. Any chat-completions model available to this deployment can be selected in Settings, and the list can be fetched live.

Does Remy Agent support multiple languages?

Yes. Choose a language in Settings to drive speech recognition and spoken output. The browser supplies the voices, so available languages depend on your device, and the model replies in the language you use.

Can I use Remy Agent on mobile?

Yes. The interface adapts to phones and tablets. Typing works in every browser; voice input needs a browser with the Web Speech API, such as Chrome or Edge on Android.

Does Remy Agent store voice conversations?

Audio is never recorded or uploaded. Transcripts are kept only in your own browser storage when conversation history is enabled, and you can delete them at any time in Settings.

How do I start a conversation?

Press Start talking and allow microphone access, or type a message in the composer. Replying aloud can be turned off in Settings.