Sift

Global ML-based fraud detection — events + scores + decisions + labels feedback. Alpha on npm — paths verified against SiftScience/sift-ruby SDK.

MIT · runs locally with your own credentials

Quickstart

Install, set the credentials, make the first call.

Pick your client. The server runs locally over stdio, with your own credentials.

Tools

What your agent can call.

Send a fraud signal to Sift's Events API (POST /v205/events). The `type` parameter selects the event kind ($create_order, $transaction, $login, $create_account, $update_account, $chargeback, $order_status, etc). Additional fields ($user_id, $session_id, $order_id, custom fields) are passed via `fields` and merged into the body. Use `return_score=true` to get a synchronous score in the response.

Fetch the latest Sift score(s) for a user (GET /v205/users/{user_id}/score). Does NOT trigger a rescore — it returns whatever score was last computed. The score is a float in [0, 1]; higher means more fraud-like. Optionally filter by abuse_types.

Force Sift to recompute a user's score right now (POST /v205/users/{user_id}/score). Use when you have externally-observed signal that should invalidate the last score (e.g. a manual decision) but have not sent a new event.

Label a user as fraud or not-fraud via the legacy Labels API (POST /v205/users/{user_id}/labels). Labels are the classic supervised-learning feedback channel for Sift's ML model. Note: most new integrations use the Decisions API (apply_decision_to_user) instead — labels are kept for backward compatibility. Still supported for ongoing model feedback.

Remove any existing label on a user (DELETE /v205/users/{user_id}/labels). Optionally scope by abuse_type.

Apply a workflow Decision to a user (POST /v3/accounts/{account_id}/users/{user_id}/decisions). Decisions are the modern replacement for Labels — they both classify the entity for Sift's ML and trigger any configured side effects (e.g. a Block decision on a $payment_abuse user will cause that user's future $transaction events to be blocked). Requires SIFT_ACCOUNT_ID.

Apply a workflow Decision to a specific order (POST /v3/accounts/{account_id}/users/{user_id}/orders/{order_id}/decisions). Order-level decisions target a single transaction rather than the whole user (e.g. 'approve_order_payment_abuse' on a manually-reviewed high-value order). Requires SIFT_ACCOUNT_ID.

Fetch the decisions currently applied to a user (GET /v3/accounts/{account_id}/users/{user_id}/decisions). Returns the active decision per abuse type. Requires SIFT_ACCOUNT_ID.

Fetch the decisions currently applied to an order (GET /v3/accounts/{account_id}/orders/{order_id}/decisions). Note: the order-level GET path does NOT include user_id, unlike the order-level apply path. Requires SIFT_ACCOUNT_ID.

Fetch the status of a Sift Workflow run (GET /v3/accounts/{account_id}/workflows/runs/{run_id}). The run_id is returned in the response to events sent with return_workflow_status=true. The run status tells you which route of the workflow was taken and which decisions were applied. Requires SIFT_ACCOUNT_ID.

Send a $chargeback event to Sift's Events API (POST /v205/events). Convenience wrapper around send_event for the most common ML-feedback signal: a confirmed chargeback. Sift uses chargeback events as a strong negative-label signal for the payment_abuse model. Optionally request a synchronous score with return_score=true.

Send a $login event to Sift's Events API (POST /v205/events). Convenience wrapper for tracking login attempts — both successful and failed attempts feed Sift's account_abuse model.

Compatibility

MCP is a protocol. Any client that speaks it can mount this server.

codespar_discover lists this server's tools, but a project with a test key cannot call them yet: the runtime has not recorded what this provider does in test mode, and refuses rather than guess.

Start in seconds.

Add the config above, set the variables, and your agent calls the tools.