Docs.
Everything here is mirrored in markdown — append .md to any URL or send Accept: text/markdown. Index for agents: /llms.txt
# remote — Streamable HTTP, OAuth 2.1\nhttps://hunter-seeker.io/api/mcpQuickstart
Target: first successful call in under five minutes. No signup needed for the demo tenant.
# 1 · connect (remote MCP)\nhttps://hunter-seeker.io/api/mcp# 2 · paste into your agent\nConnect to the Hunter-Seeker MCP server, call hs_describe_capabilities,\nthen rank the demo churn dataset by likelihood of churn and return the\ntop 10 with evidence and provenance.Free tier: the full MCP tool surface, 3 runs/month on your own data, unlimited demo tenant, and the sample datasets forever. Per-client install for Cursor, Claude, VS Code, LangChain and CrewAI is on the agents page.
MCP server
Fourteen tools — one costs a run, two cost a decision, eleven are free — with at most eight natural-language parameters each, typed errors, idempotency, and the trust contract in every envelope. Eight of them analyse a population; six decide about one entity and prove it afterwards. The server fronts the product layer only, never the engine.
Input contract, problem shapes, trust guarantees, and worked multi-domain examples. Call first when unsure.
Register tabular data past the inline cap. Presigned upload or guarded https fetch (public hosts only). direct_upload streams a large dataset (up to ~1M rows) straight to storage, past the proxy body cap.
Rank rows by likelihood of a yes/no outcome; top-k with calibrated scores, top_factors, lift, gate verdicts, leak-guard, provenance.
Poll a pending async run. Returns pending — optionally with a leak-firewalled stage and facts_so_far plus a live status_url — or the completed ranking envelope.
Minimal feature changes associated with exiting the risk pattern, for entities in a prior ranking. Rate-capped per entity.
The pattern the engine found for a prior ranking: a combination of conditions (feature + direction + threshold) with prevalence and lift.
Lift, calibration coefficient, gate verdicts, and drift history for a prior ranking.
Context metrics and exposure/recourse figures attached to a prior ranking.
Decision time. Score one entity against the scorecard a cleared ranking already fitted. Returns a calibrated score, a band — act, escalate, or refuse, with the reason — an autonomy ceiling, up to four principal reasons, and a signed Verdict.
The same decision, for many entities in one call against one model_ref.
Check a Verdict’s signature against the published keys, so a reviewer who does not trust you can confirm a decision independently. No account needed over REST (POST /v1/verify-verdict) or with the published verifier libraries, which check a signature offline with no server at all; over MCP it takes a credential like every other tool, because MCP authenticates the connection rather than the call.
Record what actually happened. Append-only: a fact store never revises history.
Record that you acted on a lever, and how far. Compliance is evaluated inside the engine; only a hash of the lever token is stored.
Whether acting is working: outcome rates for entities acted on versus not, with a confidence interval. Returns nothing below the floor — a difference computed on four entities is not evidence.
REST API
Every tool, over plain HTTP, for clients that don’t speak MCP — one operation per tool at /api/v1/, same handlers, same answers. OpenAPI 3.1 schema at /docs/openapi.json; contract 1.4.0 is served in parallel.
POST https://hunter-seeker.io/api/v1/scores\n\n{ "outcome_column": "churned",\n "entity_column": "account_id",\n "horizon": "90d",\n "page": { "k": 20 } }Model quality
Every ranking carries its own quality block: top-decile lift on out-of-time holdout, the calibration coefficient refit that cycle, the gate verdicts with their thresholds, and drift history for the series. Call hs_model_quality with a ranking_ref to re-read it without spending a run.
Governance export
One bundle per run: gate decisions with recorded reasons, leak-guard exclusions and why, model quality, drift history, the append-only action ledger, the data fingerprint with ordered run events, and a bundle sha256. Shaped for SR 26-2 model-risk documentation and EU AI Act review.
Honest-empty
Below a validated lift of 1.5 on out-of-time holdout, the engine returns a structured refusal rather than a weak ranking. The refusal names the best candidate, the bar it missed, and the failing gate. Refused runs are never billed.
{ "result": "none",\n "reasons": ["validated lift 1.31 < 1.5 surface bar"],\n "gate_verdicts": ["holdout gate: FAIL"],\n "provenance": { "engine_version": "0.3.5",\n "core_hash": "4d24e9fc…" } }Security & data
Raw data is processed for the run and then discarded; we retain derived facts and entity IDs only. Per-tenant isolation with row-level security, no bulk export beyond your own entities, and per-tenant rate caps on counterfactual access. We do not train on your data.