AI chatbots

Companions that know where the user is.

Voice and video companions already hear what the user says. With the user’s consent, orgasm.dev tells the model how close they are and what has just happened, so the companion can pace the conversation, change register at the right moment and know when it is time for aftercare.

  1. 01

    The user’s camera

    The video call they already make, published to the enclave

  2. 02

    orgasm.dev

    Imminence every second; each orgasm’s onset, peak and end

  3. 03

    Companion state

    A phase and a count, folded from the events

  4. 04

    Your model

    Reads one line of it before every turn

Signals

What a companion reads

imminence.updated

p30 and p60, about once a second: the probability that orgasm begins within 30 and within 60 seconds.

Pacing: far off, building, near the edge.

orgasm.onsetorgasm.peak

An orgasm has begun (women) or peaked (men), called a few seconds after it did.

A brief acknowledgement, and no new instructions.

orgasm.end

When it was over.

Aftercare: a slower, warmer register.

orgasmIndex

The session’s running count, carried on every call.

Something the companion can refer back to.

confidencenull

How complete the input was; no reading while calibrating or with no face in view.

When unsure, the companion says nothing about it.

Every field is described in signals; the edge, ruin and finish presets are reference policies to copy and tune.

What you can build

A companion that reads the moment

Pacing

The companion takes its time while the user is far off, and slows down and shortens its turns as they near the edge.

imminence.updated

Aftercare

Once the end is called, a slower, warmer register: it checks in, suggests water and rest, and gives no new instructions.

orgasm.end

Personas on the presets

A teasing persona that runs the edge preset, or a denial persona on ruin, with the rules and a safeword agreed before the session starts.

edge presetruin preset

A count it can refer to

The session’s orgasm count in the model’s context, so the companion can acknowledge the second or the third, and forget it when the session ends.

orgasmIndex

A toy alongside

The companion drives the user’s toy on the same cues it paces itself by, and reports each change. The sex toy case study shows the loop.

reportStimulation

Agents with tools

A companion built as an agent follows the session itself through the action MCP server: the latest imminence before each turn, the calls as they come.

get_latest_imminencewait_for_event

On the sandbox

What the model reads, turn by turn

The sandbox’s man-peak scenario, folded into the reference companion state. The model never sees the numbers: over thirteen minutes it reads six different lines, each a phase and the count.

A man's orgasm: peak, then end: imminence and orgasm calls over 13 minutes00.510 min2 min4 min6 min8 min10 min12 minbaselinesubject: manpeakend
A man's orgasm: peak, then end. The face reads as a man’s, imminence climbs to a peak, which is called about three seconds after it happens; the end is then read from his face. Bold line: p30; light line: p60. Solid verticals: the moment a call names; dashed: when it was made.
atthe model readsbecause
0:00User state: not available yet.the session starts; imminence follows the baseline
1:02User state (advisory): far off; orgasms this session: 0.p60 0.02
6:07User state (advisory): building; orgasms this session: 0.p60 0.20
8:40User state (advisory): near the edge; orgasms this session: 0.p60 0.50
11:44User state (advisory): orgasm; orgasms this session: 1.peak called, 3.2 s after it happened
12:25User state (advisory): aftercare; orgasms this session: 1.end called

Integration

The camera you already have, one line of context

The app already holds the user’s camera for the video call. Publish it from the user’s device straight to the enclave, not through your servers, so nobody sees the video. Fold each event into a compact state on your server, and put one line of it in the model’s context before every turn.

companion-state.ts
import { threshold } from './hysteresis'; // from orgasm.dev/docs/concepts/timing

const building = threshold(0.2, 0.1);
const near = threshold(0.5, 0.4);
const state = { phase: 'calibrating', orgasms: 0, aftercareUntilS: 0 };

// Fold every event into the state, in sequence order.
export function fold(event) {
  const t = event.streamTimeS;
  switch (event.type) {
    case 'imminence.updated': {
      const { p60 } = event.data;
      if (p60 === null || state.phase === 'orgasm' || t < state.aftercareUntilS) return;
      const isNear = near(p60, t) === 'above';
      const isBuilding = building(p60, t) === 'above';
      state.phase = isNear ? 'near the edge' : isBuilding ? 'building' : 'far off';
      return;
    }
    case 'orgasm.onset':
    case 'orgasm.peak':
      state.phase = 'orgasm';
      state.orgasms = event.data.orgasmIndex;
      return;
    case 'orgasm.end':
      state.phase = 'aftercare';
      state.aftercareUntilS = t + 180;
      return;
  }
}

// One line before each turn: a phase and a count, never the raw numbers.
export const userState = () =>
  state.phase === 'calibrating'
    ? 'User state: not available yet.'
    : `User state (advisory): ${state.phase}; orgasms this session: ${state.orgasms}.`;
system-prompt.md
Before each turn you see one line, "User state", from orgasm.dev. It estimates from the
user's camera how close they are to orgasm. It is advisory and sometimes wrong.

- far off, building: keep your pace.
- near the edge: slow down and keep your turns short; follow the rules the user chose.
- orgasm: a brief, warm acknowledgement; no new instructions.
- aftercare: a slower, warmer register; check in on how they feel.
- not available: say nothing about it.

Never present the state as fact, and never mention the numbers unless asked. Stop at once
when the user says stop or their safeword, whatever the state says.

Safety

What every integration builds in

Adults who said yes

Everyone on camera is an adult you verified, who agreed to this analysis before it started. A self-declared birthday is not verification.

Never covertly

The person sees when analysis is running and can end it at any moment. Consent to one experience is not consent to another.

A stop that always works

One control ends stimulation or the experience at once, without depending on the signals, the network or your server.

Signals advise; limits decide

Never raise intensity or take an irreversible step on a signal alone. Cap intensity in the device, and treat a null reading as unknown and hold.

False calls happen

Detectors miss real orgasms and call ones that did not happen. Any consequence of a call is agreed in advance and can be contested.

Keep as little as you can

The events are data about a person’s sex life. Keep them only as long as the experience needs, down to none, never sell them, and delete them on request.

And in a companion

  • The stop lives outside the model. Enforce stop and the safeword in your app as well as in the prompt; a model can miss an instruction, and the app must not.

  • Analysis is never silent. An indicator shows for as long as analysis runs, and the user can turn it off without ending the conversation.

  • Not a gender. The subject read chooses detectors. Never let the model address or describe the user by it.

  • Not a memory. Keep the events out of long-term memory and training data unless the user agreed to that specifically.

  • No upsell at the edge. Never time an upgrade, a tip or a paywall to the user’s state.

Read the acceptable use policy, the presets’ safety notes and how the enclave keeps the video private.

Other industries

The same signals, elsewhere