Pulse is a wearable and phone app that streams a speaker's live transcript and heart rate into an MCP server, so any connected AI client can read stress signals during a talk and coach the speaker in real time, not just after the fact.
What this enables
- Streams a live transcript and heart rate reading into MCP resources during a speech
- Lets any MCP client pull session vitals and generate a stress timeline on demand
- Sends short coaching prompts to an earbud through a co-pilot action mid talk
- Flags unusually high stress and prompts the speaker to pause and breathe
The signal AI agents were missing
AI agents can already read a person's files, calendar, code, and email. What they cannot read is the person delivering the pitch, the interview answer, or the presentation. A transcript alone tells an assistant what was said. It says nothing about whether the speaker was calm, rushed, or spiking with adrenaline three minutes in. That gap is what Pulse was built to close.
Why a live speech is a hard input to work with
A speech is not a document an agent can fetch once and reason over. It is a moving target: words are still being spoken, stress is changing second to second, and any useful feedback has to arrive while the talk is still happening, not in a report generated afterward. Solving that meant giving an AI client two things it does not normally have: a transcript that updates as the speaker talks, and a physiological signal running alongside it, both readable without the speaker doing anything more than tapping a watch.
How Pulse is built
The system has three parts: a watch app, a phone companion app, and an MCP server. Tapping the watch starts a session. The phone begins transcribing the speech in real time, while the watch continuously streams heart rate and other vitals. The MCP server exposes this as two resources: a live session transcript and the current session vitals, both readable by any connected client.
From there, an assistant like Claude or Copilot can query those resources directly. Asking for a session report pulls the transcript and heart rate stream together into a heart rate sync graph, showing how the speaker's pulse moved against what they were saying at each point, for example opening around 95 bpm and settling into the high 80s and high 70s as the talk went on.
The same connection supports a co-pilot action: pressing it mid speech returns a short spoken prompt, delivered to an earbud, based on what the live data shows at that moment.
Where NitroStack fits
The transcript and vitals are exposed as MCP resources built on the NitroStack SDK, which is what lets an assistant pull fresh session data on demand instead of waiting for a one time export. The session report itself renders through a NitroStack widget rather than as plain text, which is why the heart rate sync graph shows up as an interactive chart tied to transcript segments instead of a wall of numbers.
Why resources instead of tools here
The interesting design choice is treating the transcript and vitals as resources, not tools. A tool implies an action being taken. What the assistant actually needs here is continuous read access to state that is changing on its own, which is what a resource is for. The trade off is that the whole loop depends on the watch staying connected to the phone. If that link drops mid session, the vitals resource goes stale and the assistant starts reasoning from old data without any obvious signal that it has done so. Real time coaching prompts carry a similar risk: useful when timed well, distracting if they fire too often during a live talk.
The reusable pattern
The pattern worth taking from this is not "AI can read your heart rate." It is that a live biometric or behavioral stream can be exposed as a readable MCP resource, separate from any action the assistant takes, so multiple clients can ask "how is this person doing right now" without each one needing its own bespoke integration into the wearable.
If you want to see Pulse in action, check out the complete video here:
It walks through tapping the watch to start a session, how the phone streams the live transcript, how the MCP resources expose heart rate and transcript data, and how the co-pilot action delivers a real time coaching prompt to an earbud.
To explore the implementation in more detail, head over to the NitroStack docs for guides on building tools, resources, prompts, testing MCP servers in NitroStudio, and deploying them through NitroCloud.
Check out the complete code over here. You can also read the source and build your own MCP application using the same SDK and workflow at nitrostack.ai.