Finic's Kairos: open-source fraud decisions that a coding agent tunes

Finic, a start-up, has released Kairos, an open-source platform that decides in real time whether a sign-up, login, order or payout should go through. It appeared on Hacker News as a Show HN on 3 November 2026 and is licensed under Apache 2.0, according to its GitHub README. Finic positions it as an alternative to Sift, Sardine, Riskified and SEON. Everything below is Finic's own description; the post had 13 points and no comments when I read it.

What it does

For each request, Kairos returns approve, step up, review or block, with a risk score and reasons. Its docs say it reads the subject's last 2,000 events, builds a context of the 30 most recent events in full plus a statistical summary of older ones, and checks how many other subjects share the same device or payee. It then asks Finic's model, Grayson, a yes/no question that you write in plain language, and applies a policy of thresholds and rules. The docs give a default of blocking at 0.9 and reviewing at 0.5 and call these placeholders.

Finic's pitch is that this works before a company has the thousands of labelled chargebacks a traditional model needs. Where labels exist, Kairos can also train logistic regression and XGBoost-compatible trees, with Grayson's answers as extra features.

The agent-as-analyst part

The distinctive claim is that a coding agent does the work of a fraud machine-learning engineer. Through an MCP server, Claude Code, Cursor or Codex can define workflows and propose model variants. Kairos trains each on the oldest labelled data, selects on the next slice, tests on the newest, and reports ROC-AUC, PR-AUC and the gain over a baseline with a 95% interval. The README's example has an agent publishing a variant with 0.949 ROC-AUC on validation and 0.930 on test; that is an illustration in the README, not a published benchmark.

What it runs on, and what it leaves out

Self-hosting is one Node.js 22 process plus Postgres, run with Docker Compose. The model is not self-hosted: the server calls the Finic API and needs a Grayson key, so decisions leave your network and the per-decision price is not stated in the README or the overview docs I read.

The docs are candid about limits. In a test of 100 repeated requests, 25 returned a different probability, by at most 0.06, so a score near a threshold can land on either side. There is no shadow mode and no threshold tuning, step-up flows are left to you, and links are followed one hop only. No fraud-catch or false-positive rates are published.

What to do with it

If you have no fraud tooling and few labels, run Kairos against historical events for which you know the outcome, and compare its review queue with what your team actually caught. Check the cost of the hosted model at your volume, and decide who approves a change before an agent is allowed to publish a new workflow version.