This is a submission for the Presage SmartSpectra SDK Challenge.
What I Built
AI assistants only get your words. They can't tell when you've gone quiet after a bad day or when your pulse is climbing during a tense conversation. People pick up on these cues all the time. AI assistants miss them.
Attune gives an AI a sense of how you're doing, with your consent, using only a regular webcam. It runs Presage SmartSpectra on your own computer. It turns the live readings into short guidance for the AI, never raw numbers. For example:
"They seem a little tense. Slow down, check in tentatively and invite correction."
You can use Attune in two ways:
- Built-in Companion. A chat (and optional live voice) assistant inside the Attune Windows app. Its tone changes with how you seem to be doing.
-
A local MCP connector. Claude Desktop, ChatGPT (through Secure MCP Tunnel) or any stdio MCP client can call
get_body_contextbefore replying. Each app only gets the "lenses" you approve, such as companion, yoga/breathwork or casual chat.
The video covers:
- Consent first: camera sensing, AI sharing and memory are separate opt-ins.
- Companion: the user says they made a mistake at work. Attune picks up a quieter, lower-energy state and the reply becomes gentler. The panel above the chat shows the exact guidance the AI received for that turn.
- Claude Desktop over MCP: Claude asks Attune for body context, the user approves the purpose, and Claude reads the guidance before it answers.
- Audit trail: every read is logged with the app, lens, time and state, and never the numbers.
Code
GitHub Repository: clatsonhacks/attune
Setup (Windows, Node.js 26):
git clone https://github.com/clatsonhacks/attune.git
cd attune
npm ci
npm start
On first launch, API Keys lets you add a Presage SmartSpectra key (camera sensing), a Groq key (built-in chat) and an OpenAI key (live voice). Every key is optional and stored encrypted for your Windows user.
Want to try it without keys or a camera? Run:
npm run desktop:demo
The README also covers the web-only service, the MCP connector and building the Windows installer.
How It Responds to Human Signals
Signals used: pulse rate, HRV, breathing rate, facial expression, talking detection and blinking, plus Presage Insights check-ins.
Integration
@smartspectra/node-sdk runs in its own worker process (src/adapter/worker.js), so the camera and native runtime are isolated from the rest of the app.
src/adapter/smartspectra.js creates a SmartSpectraSDK on the default webcam with enableTelemetry: false.
It requests the cardio, breathing and face metric sets and maps each decoded Metrics packet to a simple sample shape (cardio.pulseRate, cardio.hrv, breathing.rate, face.expression, face.talking, with confidence normalized to 0–1).
If your chest isn't in view for 20 seconds, it falls back to face-only metrics.
From Readings to Behavior
The interpretation logic lives in src/interpreter.js.
Personal baseline. The first ~20 seconds calibrate your calm baseline. Signals are judged against your own baseline, not population averages, which also handles webcam noise.
Sustained and hysteresis. A rise of about 6 BPM over baseline has to last 8 seconds before Attune nudges the AI. It has to settle for 6 seconds before it clears, so the AI doesn't keep flip-flopping.
Confidence gating. Below 0.4 confidence the state becomes
unclear, and the AI is told to say nothing about your body.Corrections. If you say "no, I'm fine",
record_correctionraises your threshold, up to a cap. Attune learns how you show stress.Lenses. The same reading becomes different guidance for different purposes. In yoga it's "Gently invite a longer exhale", and in casual chat it's "Keep it light and follow their lead."
Insights. Presage check-ins arrive automatically, or on request with Ask for a check-in, limited to once every 30 seconds. You see the full text. The AI gets a short note with numbers and medical wording removed.
What the AI Receives
The AI receives a structured context object:
{
"state": "tense",
"say": true,
"directive": "Slow down and check in tentatively.",
"suggestedUtterance": "Would you like to take a moment?",
"confidence": 0.85
}
This context is refreshed every turn and never blocks on the camera.
In live voice (OpenAI Realtime over WebRTC), the current pacing cue goes over the call's data channel, so the spoken reply adjusts mid-conversation too.
Prize Categories
Best Thoughtful Real-World Use, and also Most Creative or Delightful Use.
Real-World Need
More and more people talk to AI about stressful things: work mistakes, anxiety, loneliness. An assistant that can't notice when someone is overwhelmed answers everyone in the same upbeat tone.
Attune is for anyone who uses an AI companion, and for developers who want their assistant to be more emotionally aware without building sensing themselves.
It's designed to be usable:
- No wearable, just the webcam already on your laptop.
- Consent is asked once in plain language.
- A visible Camera on badge shows when sensing runs.
- On-screen tips help with positioning.
- Pause/forget is one click away.
The AI is told to check in tentatively and ask for correction, never to state conclusions about your body.
What's Original
Attune isn't another vitals dashboard.
It's a body-sense layer for any AI.
Because it's an MCP server, Claude, ChatGPT or your own agent can become more attentive without touching the camera, the SDK or a single heartbeat value.
Consent and Data Handling
Explicit, Layered Opt-In
Camera sensing (with the Presage disclosure), AI sharing and memory are three separate choices.
Nothing starts until you agree.
The camera starts only when you turn it on, start a chat, or when an app you approved asks.
It turns off after 5 minutes idle.
Pause stays on until you resume.
Forget camera permission deletes the saved consent.
No AI can grant a permission, and the connector has no consent, approval or forget tools.
Per-App, Per-Purpose Approvals
Each connected app reads only through the lenses you allow in Attune → Connectors.
ChatGPT gets nothing until you approve it.
Unapproved requests never start the camera.
What Leaves Your Device
Camera frames go only to the SmartSpectra SDK in a local worker process.
The SDK talks to Presage services for sign-in and Insights, which can include physiological measurements, and this is disclosed in the consent dialog.
SDK telemetry is turned off.
AI apps receive only predefined guidance and state.
Pulse, HRV, breathing values, raw Insight text and video never go to an AI.
Optional AI Services
Built-in chat messages (last 20 turns, kept in memory only) go to Groq when you add a Groq key.
Live voice streams microphone audio directly to OpenAI over WebRTC with your own key, only while a call is active.
Claude and ChatGPT use their own accounts.
Storage and Retention
Consent, approvals, an activity log (app, lens, time, state, never numbers) and optional number-free memory notes are stored locally in %APPDATA%\Attune.
You can clear the log and memory from the app.
API keys are encrypted for your Windows user.
The local API accepts only 127.0.0.1 with a random token and rejects browser-origin requests.
Not Medical
Attune makes no diagnosis or treatment claims and isn't for medical decisions.
It only shapes conversational tone.
No covert monitoring: sensing is always visible and in your control.

Top comments (0)