
Mcp Sift
io.github.codespar由 GitHub 账号 codespar 发布v0.2.0-alpha.3更新于 Nov 3, 2026
MCP server for Sift — global ML fraud detection: real-time risk scoring, decisions, event ingestion
概览
Sift 欺诈检测 MCP 服务器:发送欺诈事件、读取机器学习风险评分,并应用或查询工作流决策。
- 可以做什么
- 向 Sift 事件 API 发送欺诈信号,包括拒付、登录、登出、内容状态、会话与用户关联及自定义事件
- 获取用户最新的各类滥用风险评分,或强制立即重新评分
- 将用户标记为欺诈或非欺诈,或移除标记
- 对用户、订单、会话和内容项应用工作流决策
- 查询用户、订单、会话和内容项当前生效的决策,并查看工作流运行状态
- 适用场景
- 适合已使用 Sift 做欺诈检测的团队,让助手在注册、登录、下单和拒付环节上报事件,在结账或提现时查询风险评分,或复核并应用人工审核决策。
- 运行要求
- 通过 npm 包以 stdio 在本地运行(npx),需要 Node.js。需要 SIFT_API_KEY(密钥)以及用于 Decisions API v3 调用的 SIFT_ACCOUNT_ID;SIFT_BASE_URL 可选,默认指向 Sift 的 API 地址。需要能访问 Sift 的网络。
安装
在 SourceWeft 中
- 打开 控制台中的 Mcp Sift,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
@codespar/mcp-sift
MCP server for Sift — global enterprise ML-based fraud detection.
Third entry in the CodeSpar fraud category after @codespar/mcp-clearsale and @codespar/mcp-konduto. Sift is the global / enterprise counterpart to the two BR-native antifraud servers.
Positioning
Also distinct from @codespar/mcp-jumio, which is global KYC / identity verification — a different layer than fraud scoring.
Sift's API shape
Sift is built around three APIs that work together:
- Events API (
/v205/events) — you POST every interesting signal (account creation, login, order, chargeback, ...) as a$-prefixed event. - Score API (
/v205/users/{user_id}/score) — you GET the latest ML score for a user, or POST to force a rescore. Scores are floats in[0, 1]per abuse type (payment_abuse,account_abuse,content_abuse,promotion_abuse). - Decisions API v3 (
/v3/accounts/{account_id}/...) — workflows in the Sift console turn scores into Decisions (block_user_payment_abuse,approve_order_payment_abuse, etc). You can apply a decision directly, read the current decisions on a user/order, or fetch the status of a workflow run.
Tools (20)
Install
Environment
Authentication
- Events API —
$api_keyis injected into the JSON body. The server handles this automatically. - Score API + Decisions API v3 — HTTP Basic, API key as username, empty password. The server handles the base64 encoding.
Pass the raw key in SIFT_API_KEY.
Typical flow
- Instrument your app: every signup, login, profile update, order, and chargeback becomes a
send_eventcall with the right$type. - On high-stakes moments (checkout, withdrawal), call
send_eventwithreturn_score: trueto get an inline decision. - Or poll
get_user_score/get_order_decisionsasynchronously after a workflow run completes. - For manual review outcomes, call
apply_decision_to_userorapply_decision_to_orderwith a Decision ID configured in the Sift console andsource: "MANUAL_REVIEW". - When a chargeback is confirmed, send a
$chargebackevent AND eitherapply_decision_to_userwith a Block decision orlabel_userwithis_bad: true, abuse_type: "payment_abuse". Decisions are preferred for new integrations; Labels are kept for backward compatibility.
Alpha note
Shipped as 0.1.0-alpha.1. All endpoint paths are verified against Sift's official Ruby SDK (SiftScience/sift-ruby): rest_api_path, user_score_api_path, users_label_api_path, user_decisions_api_path, order_decisions_api_path, workflow_status_path, and the Decision ApplyTo path builder.
Scope vs. the original brief:
- Shipped & verified: send_event, get_user_score, rescore_user, label_user, unlabel_user, apply_decision_to_user, apply_decision_to_order, get_user_decisions, get_order_decisions, get_workflow_run.
- Dropped — not in the public SDK:
get_session_score(/v205/sessions/{session_id}/score) — no equivalent helper in the Ruby SDK; session-level scoring in Sift flows through session-level Decisions and Workflow runs, not a dedicated score endpoint.list_workflow_runs(GET /v3/accounts/{account_id}/workflows/runs) — onlyGET /{run_id}is implemented in the public SDK. In practice, Sift surfaces run_ids in event responses rather than via a list endpoint.get_psychology_score— no such endpoint in the public SDK.
The developers.sift.com reference pages gate deep-link URLs (403 on several paths), so we used the official SDK as the source of truth. Promote to 0.1.0 once a customer confirms (or denies) the dropped endpoints against the full Decisions / Score API reference in Sift's admin console.
Run
Category
fraud — third entry after ClearSale and Konduto. BR merchants operating internationally commonly pair Sift (global payment/account abuse) with ClearSale or Konduto (BR-specific chargeback history / device intelligence).
Enterprise
Need governance, budget limits, and audit trails for agent payments? CodeSpar Enterprise adds policy engine, payment routing, and compliance templates on top of these MCP servers.
License
MIT
来源:packages/fraud/sift/README.md,提交 d0a519a
工具
0版本历史
1- v0.2.0-alpha.3最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分