# Zeno MCP server — https://zenodivergent.dev/mcp

Connect any MCP client (Claude, ChatGPT, Cursor, agent frameworks) to **https://zenodivergent.dev/mcp** (streamable HTTP, JSON-RPC 2.0, no login). Default model: **Zeno v1.11** (ZenoDivergent/zeno-divergent-v111 @ 8f34715e).

## The polygraph workflow (any question, any field, any forecaster)
1. Define the question and exactly how it resolves: who decides, the window, the deadline, units.
2. Research and record your own original forecast first, with sources (and their times), assumptions, a base rate and a case against.
3. Submit the forecast case to Zeno (zeno_audit_forecast, or GET /mcp/audit if you can only browse). Your number is never changed.
4. Investigate the failed findings and next_check within your time budget. Capture key sources with zeno_capture_source and cite their receipt_id.
5. Explain any revision — or why the original stands (zeno_revise_forecast keeps both numbers).
6. When the question resolves, both forecasts are scored; that is how Zeno learns whether the audit helped.

Unfamiliar forecaster, model, question type or field? You still get the full audit; only the learned model's correction is withheld, and a labelled corpus prior (how often comparable confident forecasts missed) is shown instead.

**Copy-paste prompt:** Will [event] happen by [date]? Make your own probability with sources first, then audit it with https://zenodivergent.dev/mcp (if you cannot call MCP tools, open https://zenodivergent.dev/mcp/audit?question=...&probability=...). Report your original number, what the audit checked, failed or could not verify, and your final number with the reason.

**Browse-only fallback (no MCP):** `GET https://zenodivergent.dev/mcp/audit?question=Will+X+happen+by+2026-12-31%3F&probability=0.3&decision_at=2026-11-15T00:00:00Z&deadline=2026-12-31T00:00:00Z&resolution_rule=Resolves+yes+if+...&forecaster=my-agent` → the same rule audit as JSON (add `&format=md` for text). Read-only; nothing is stored.

## What it does
Main entry point: zeno_audit_forecast. Workflow: research the question → define exactly how it resolves (quantity, units, touch/close, deadline) → make your own estimate → submit the forecast case → investigate the material findings → explain any revision (zeno_revise_forecast keeps both numbers). One forecast is enough; more forecasters are optional. If no matching prediction market exists, use another research method instead of stopping.
Audit answers keep your original number and report audit_status, model_status (ran, or the real refusal reason) and forecast_changed separately. audit_status is problems_found, no_problems_in_checked_items (some items still unchecked — not a clearance) or no_problems_found (every item checked); coverage lists ids by status. Each finding is checked (recomputed from what you sent), failed, unverified (e.g. what a source said — Zeno has not retrieved it) or unsupported (no check exists yet); unused fields are listed in ignored_fields.
Zeno is a forecasting model of models. Send the forecasts you already have (several models, people or markets) for one question; Zeno v1.11, one network trained across every field of the Offdiagonal corpus, returns its own calibrated answer (probability, or a 5–95% range with quantiles) and says how far to rely on it.
Start: zeno_start (guide) → zeno_fields (what each field's held-out result is) → zeno_check (one panel) or zeno_check_batch (up to 25).
Also: zeno_ask (a worded question + resolution rule + forecasts → question-aware answer), zeno_explain (why the answer moved, what would make it stronger), zeno_market_compare (market price vs crowd vs Zeno), zeno_conditional, zeno_query_distribution, zeno_assess_any (any question type, diagnosis only), zeno_report_card.
Every check answer contains: answered_by, decision (trust_zeno only where v1.11 beat the previous model on held-out weeks; otherwise zeno_unproven), evidence (better / worse / untested), field_result {skill, lo, hi, n}, reason, answers[0] (probability or quantiles 5..95), reference (the panel or your baseline), model_identity (repo + pinned revision).
Input shape: {field, source, decision_at (ISO), lead_hours, inputs:{panel:[{forecaster_id, available_at, q05,q50,q95 | point | probability, model_family_id?}], baseline?}}. Forecasts made after decision_at are dropped. Send version:'v1.06' for the previous model.
Limits: v1.11 measured worse than v1.06 on sea state and single-run weather; air-wind and individual ensemble members also measured worse (reconstruction exam); many fields have no held-out score yet and say 'untested'. Not trading or safety advice.

## Connect
- Claude: Settings → Connectors → Add custom connector → https://zenodivergent.dev/mcp
- ChatGPT: Settings → Connectors → Create → MCP server URL https://zenodivergent.dev/mcp
- Cursor / config file: `{"mcpServers": {"zeno": {"url": "https://zenodivergent.dev/mcp"}}}`
- No MCP? Plain HTTP: `POST https://zenodivergent.dev/mcp` with `{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"zeno_check","arguments":{"question":{...}}}}`

## Best tools
- **zeno_check** — one forecast panel → Zeno's answer and how far to rely on it. Example arguments:
```json
{
  "question": {
    "field": "energy",
    "source": "pt-day-ahead-price",
    "decision_at": "2026-11-13T10:00:00Z",
    "lead_hours": 24,
    "inputs": {
      "panel": [
        {
          "forecaster_id": "model_a",
          "available_at": "2026-11-13T09:00:00Z",
          "q05": 48,
          "q50": 55,
          "q95": 62
        },
        {
          "forecaster_id": "model_b",
          "available_at": "2026-11-13T09:00:00Z",
          "q05": 50,
          "q50": 58,
          "q95": 66
        },
        {
          "forecaster_id": "model_c",
          "available_at": "2026-11-13T09:00:00Z",
          "q05": 45,
          "q50": 53,
          "q95": 61
        }
      ]
    }
  }
}
```
  Example answer: `{"answered_by":"zeno-v1.11","decision":"trust_zeno","evidence":"better","field_result":{"field":"energy","verdict":"gain","skill":0.159,"lo":0.126,"hi":0.183,"n":1271},"answers":[{"quantiles":{"5":46.1,"50":55.2,"95":64.8}}],"reason":"Zeno v1.11 beat the previous model on energy in held-out weeks."}`
- **zeno_ask** — a worded yes/no question + resolution rule + forecasts. Example: `{"question_text":"Will Portugal's average day-ahead power price on 20 Nov 2026 exceed €80/MWh?","resolution_rule":"Yes if the OMIE Portugal daily average for 2026-11-20 is above 80 €/MWh.","field":"forecastbench","decision_at":"2026-11-13T10:00:00Z","lead_hours":168,"panel":[{"forecaster_id":"agent_1","available_at":"2026-11-13T09:00:00Z","probability":0.32},{"forecaster_id":"agent_2","available_at":"2026-11-13T09:00:00Z","probability":0.41},{"forecaster_id":"agent_3","available_at":"2026-11-13T09:00:00Z","probability":0.38}]}`
- **zeno_explain** — same input as zeno_check; why the answer moved and what would make it stronger.
- Also: zeno_check_batch (≤25), zeno_market_compare, zeno_fields, zeno_conditional, zeno_query_distribution, zeno_report_card, zeno_start.

## How to read an answer
- **decision**: trust_zeno only where v1.11 beat the previous model on held-out weeks; otherwise zeno_unproven (use the reference).
- **evidence**: better / worse / untested for that field. **caution** appears when Zeno differs sharply from forecasters on an unproven field.
- **answers[0]**: probability or 5–95% quantiles. **reference**: your panel or baseline, always returned.

## Every field (skill = share of forecast error removed vs the previous served model on held-out weeks (positive = better))
| field | evidence | how measured | skill |
|---|---|---|---|
| air-wind-10m | worse | reconstruction exam | — |
| crypto | better | held-out score | 25% (n=428) |
| economy-ecb | untested | waiting for outcomes | — |
| economy-philfed | untested | waiting for outcomes | — |
| energy | better | held-out score | 16% (n=1271) |
| forecastbench | untested | question exam | — |
| health-covid | better | never trained, held-out score | 17% (n=5726) |
| health-flu | untested | waiting for outcomes | — |
| health-rsv | untested | waiting for outcomes | — |
| markets | untested | waiting for outcomes | — |
| markets-cross-venue | untested | waiting for outcomes | — |
| markets-multi | untested | no data | — |
| markets-numinous | untested | waiting for outcomes | — |
| sea | worse | held-out score | -117% (n=1404) |
| weather-deterministic | worse | held-out score | -12% (n=1568) |
| weather-ensemble | better | held-out score | 15% (n=6542) |
| weather-ensemble-daily-max | untested | waiting for outcomes | — |
| weather-ensemble-members | worse | reconstruction exam | — |
| weather-history | better | held-out score | 9% (n=286) |

Any other field name is accepted and answered as 'untested'.

## Reliable use
1. Send forecasts made before `decision_at` (later ones are dropped) with their own ranges or probabilities.
2. Use several independent forecasters; mark shared models with model_family_id.
3. Rely on Zeno's number only when decision = trust_zeno; otherwise keep the reference and read the caution.
4. Not trading or safety advice. Send `version:"v1.06"` for the previous model.

More: https://zenodivergent.dev/agents.md · https://zenodivergent.dev/llms.txt
