PolyCypher

A PolyCypher monograph · 2026

The prediction engine AI agents call for biology.

Claude and ChatGPT can discuss your health. They cannot predict it. PolyCypher returns calibrated, competing-risk probabilities over time — so the agent has real biology to reason with.

Diseases
13
Interface
MCP — Claude, ChatGPT, clinical copilots
Status
Shipping this spring

Example · a conversation, in full

How it reads when Claude asks the genome.

A deterministic reconstruction of an exchange with Claude. The user asks a question; Claude decides to call the polycypher tool; the tool returns a structured answer; Claude reads it back in plain English.

User

I’m 42, Scandinavian, father had a heart attack at 56. What does my genome say about my CAD risk?

Called tool

→ polycypher.get_disease_risk(disease=“cad”, horizon_years=10)

Tool response
{
  "disease": "cad",
  "horizon_years": 10,
  "risk": 0.114,
  "population_percentile": 91,
  "ci_95": [0.094, 0.137],
  "model": {
    "version": "cad-v2.1",
    "type": "competing_risk_cox",
    "primary_ancestry": "european"
  }
}
Claude

Your 10-year CAD risk comes back at 11.4% — the 91st percentile for your age/sex/ancestry. That’s meaningfully elevated; the competing-risk model puts the 95% confidence interval at 9.4–13.7%. Given your father’s event at 56, this is worth raising with your physician for earlier screening.

Model outputs are illustrative; a real call returns live values from the caller’s own genome. CAD = coronary artery disease.

Thirty years in one frame.

A calibrated probability isn’t a number — it’s a curve. Six diseases, one hypothetical reader, thirty-five years of life ahead. Hover a line to read it; change the scenario to see how the landscape bends.

An illustrative chart showing cumulative risk trajectories from age 35 to 70 for six diseases: Type 2 Diabetes, Coronary Artery Disease, Breast Cancer, Alzheimer’s, Colorectal Cancer, and Atrial Fibrillation. Three scenarios are available: population average, an elevated-risk illustrative individual, and the same individual with an early-screening intervention applied.

Illustrative trajectories. Real predictions come from the individual’s own data — with confidence intervals, primary ancestry, and the limits of calibration returned alongside each probability.

Validated on held-out data.

Trained on large biomedical cohorts and evaluated on held-out participants. Outputs are competing-risk time-to-event probabilities over explicit horizons — not static population percentiles.

Thirteen diseases live today, each at a ten-year horizon — including Type 2 Diabetes, Coronary Artery Disease, and Breast Cancer.

  • Evaluated across ancestries. Held-out participants from multiple ancestries, not a single slice.
  • Competing-risk, not static. Explicit hazards over explicit horizons — the shape an agent can reason with.
  • Every response carries its methods. Version, confidence interval, primary ancestry, and the limits of calibration — returned in the response.

An agent can discuss your health. It cannot predict it.

The great modeling work of this century was not language. It was risk. Actuarial tables, competing-risk models, polygenic scoring — a half-century of statistical invention designed to answer one question: given what we know about a person, what is likely to happen, and when?

A foundation model is the wrong shape for this work. A genome is three billion positions of signal; meaningful effects are distributed across thousands of variants with individually small weights. Competing risk requires explicit hazards over explicit horizons. None of this lives in a language model’s attention window.

So the agent needs a tool. Something that returns a calibrated probability, the horizon it applies to, the cohort it was trained on, and the limits of its own confidence — in the shape the agent can reason with.

PolyCypher is that tool. We are not building another chat. We are building the instrument the chats will call.

One engine.
Any input. Any agent.

  1. i

    Intake

    A file enters through a widget inside the user’s chat — 23andMe export, VCF, whole-genome sequence. The bytes upload to our sealed environment, not into the agent’s context.

    • live genome & sequencing
    • next laboratory panels
    • next medical history
    • after wearables, family history, lifestyle
  2. ii

    Computation

    Imputation against reference panels, polygenic aggregation across thousands of variants, competing-risk time-to-event models trained on large biomedical cohorts. See held-out performance above.

  3. iii

    Return

    Structured JSON flows back to the agent — probability, horizon, confidence, methods, primary ancestry, and the limits of calibration. See the specimen exchange above for the live shape.

The platform shifted.

  1. Jan 26 · 2026

    MCP Apps lands.

    Anthropic and OpenAI ship the app standard for agents. Third-party tools render inline, files bypass the model’s context, and the agent becomes a host for domain-specific compute.

  2. Jan 2026

    Claude Health and ChatGPT Health launch.

    Every major AI is now a health assistant — connected to labs, records, wearables. No genomics integration ships at launch. The shelf is empty.

  3. Now

    PolyCypher ships.

    The missing piece in the stack — the tool an agent calls to understand a specific person’s biology.

People. Clinicians. AI platforms.

For people
Drop a 23andMe export, VCF, or whole-genome file into the chat you already use. The agent, grounded in your biology, keeps explaining as new models ship.
For clinicians
The same engine inside the AI copilots physicians already use. Pilots with concierge practices and diagnostic teams — methods-transparent, not a black box.
For AI platforms
Remote MCP server. Tool-call interface over JSON. No retraining, no fine-tune, no data collection on your end.

Clinicians, scientists, engineers.

  1. Portrait of Gilead Kedem

    Gilead Kedem

    Co-founder · clinical & strategy

  2. Portrait of Edward Messick

    Edward Messick

    Co-founder · engineering

  3. Portrait of Scott Sauers

    Scott Sauers

    Co-founder · science

Partner with us.

Three ways to work with us.

  1. For AI platforms

    Integrate PolyCypher as a tool.

    Remote MCP server. JSON tool contract. Zero retraining, zero on-device data.

    Start an integration conversation

  2. For clinicians

    Run a concierge pilot.

    Clinicians see confidence intervals, ancestry limits, and model version — not opaque predictions.

    Request a pilot briefing

  3. For people using AI

    Connect PolyCypher to your own agent.

    One URL in your chat settings. Your own agent gains calibrated biology over your own genome file. Shipping this spring.

    Preview Settings › Connectors › polycypher.com/mcp

    Join the early-access list

None of the above? Get in touch →

Get in touch

Write to us.

Tell us what you’re working on.

Or email hello@polycypher.com.