AI News AI资讯 3h ago Updated 2h ago 更新于 2小时前 46

Show HN: Setoku – Self-hosted knowledge server for AI agents Show HN:Setoku - 面向AI代理的自托管知识服务器

Setoku is an open-source, self-hosted MCP knowledge server that enables AI agents to query company data securely without running inference on the server. It implements a "context-first" architecture where agents must retrieve canonical metric definitions and business logic before executing read-only queries. The system features human-in-the-loop approval for knowledge updates, ensuring that AI proposals do not alter the underlying data semantics without human verification. It supports "vibe dash Setoku 是一个开源的自托管 MCP 知识服务器,旨在让 AI 模型(如 Claude Code)能够安全、准确地查询企业私有数据。 核心机制包括只读数据查询、业务上下文记忆(指标定义与陷阱)以及通过自然语言生成并发布实时数据仪表盘应用。 架构采用单租户部署,数据保留在用户本地服务器,仅通过代理暴露接口,确保无推理成本且数据不离开本地。 提供严格的安全控制,包括数据库引擎层面的 SELECT-only 权限、行级限制、审计日志以及人工审批的知识更新流程。

65
Hot 热度
70
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Setoku is an open-source, self-hosted MCP knowledge server that enables AI agents to query company data securely without running inference on the server.
  • It implements a "context-first" architecture where agents must retrieve canonical metric definitions and business logic before executing read-only queries.
  • The system features human-in-the-loop approval for knowledge updates, ensuring that AI proposals do not alter the underlying data semantics without human verification.
  • It supports "vibe dashboarding," allowing users to generate live, shareable web apps from natural language prompts backed by governed, read-only database access.
  • Designed for low-cost deployment (single $5–12/mo VPS), it uses ClickHouse for the data lake and Postgres for the knowledge store, keeping production databases isolated.

Why It Matters

This solution addresses the critical challenge of grounding LLMs in proprietary enterprise data while maintaining strict security and governance boundaries. By decoupling the AI inference layer from the data layer through a standardized MCP interface, it allows organizations to leverage existing AI models without exposing sensitive credentials or risking unauthorized writes. This approach significantly reduces the barrier to entry for building secure, data-aware AI assistants within internal teams.

Technical Details

  • Architecture: A self-hosted stack comprising a gateway, a knowledge store (Postgres), and a data lake (ClickHouse). It acts as a proxy between AI clients and internal data sources, ensuring no direct database exposure.
  • MCP Tools: Provides three categories of tools: Context tools (find_context, get_metric) for semantic understanding, Read-only query tools (run_query) for data retrieval with enforced SELECT-only constraints, and App tools (publish_app) for generating live dashboards.
  • Security & Governance: Implements row-level caps, statement timeouts, and append-only audit logs. Knowledge changes proposed by AI require human approval via an admin interface, preventing prompt injection from altering business logic.
  • Integration: Works with any MCP-compatible client (Claude, Codex, etc.) and integrates via Claude Code skills for onboarding, connecting data sources, and curating knowledge.
  • Deployment: Optimized for lightweight infrastructure, capable of running on a small VPS. Uses Docker for containerization and generates HTTPS certificates automatically during bootstrap.

Industry Insight

  • Adoption of MCP Standards: As the Model Context Protocol gains traction, tools like Setoku demonstrate its utility in bridging the gap between unstructured AI interactions and structured enterprise data, encouraging broader adoption of standard interfaces.
  • Shift to Human-in-the-Loop AI Ops: The requirement for human approval on knowledge updates highlights a growing industry trend toward hybrid automation, where AI suggests but humans validate, reducing hallucination risks in critical business contexts.
  • Cost-Efficient AI Infrastructure: The ability to run complex data-aware AI applications on minimal hardware ($5–12/mo) challenges the assumption that enterprise-grade AI requires expensive cloud inference clusters, making it accessible for smaller teams and startups.

TL;DR

  • Setoku 是一个开源的自托管 MCP 知识服务器,旨在让 AI 模型(如 Claude Code)能够安全、准确地查询企业私有数据。
  • 核心机制包括只读数据查询、业务上下文记忆(指标定义与陷阱)以及通过自然语言生成并发布实时数据仪表盘应用。
  • 架构采用单租户部署,数据保留在用户本地服务器,仅通过代理暴露接口,确保无推理成本且数据不离开本地。
  • 提供严格的安全控制,包括数据库引擎层面的 SELECT-only 权限、行级限制、审计日志以及人工审批的知识更新流程。

为什么值得看

Setoku 解决了企业级 AI 应用中“数据连接”与“语义理解”脱节的关键痛点,通过 MCP 协议实现了低成本、高安全性的私有数据接入。它展示了如何将 AI 从单纯的对话工具转化为具备实时数据分析能力的业务助手,为中小企业提供了极具性价比的私有化 AI 数据解决方案。

技术解析

  • MCP 工具集设计:提供三类工具:Context 工具(find_context, get_metric)用于检索业务含义和纠正潜在错误;Read-only query 工具(run_query)通过只读镜像执行查询;App 工具(publish_app)将结果发布为可分享的 Web 应用。
  • 混合存储架构:后端结合 ClickHouse 作为数据湖(存储日志、事件及数据库镜像)和独立的 Knowledge Store(存储元数据和业务规则),通过 Cron 任务同步数据,实现读写分离。
  • 安全隔离机制:采用“人审模式”,AI 只能提议修改知识库,需管理员在 Web 控制台批准才能生效,防止提示注入攻击导致的数据污染或逻辑篡改。
  • 低资源部署:支持 Docker 容器化部署,推荐配置极低($5-12/月的 VPS),无需运行大型模型,仅作为网关和知识管理中间件,兼容任何支持 MCP 的客户端。

行业启示

  • MCP 协议的落地实践:Setoku 证明了 Model Context Protocol (MCP) 在解决异构数据源标准化接入方面的有效性,降低了 AI 应用开发的集成复杂度。
  • 人机协同的知识治理:强调“AI 提议 + 人类审批”的知识更新闭环,为构建可信的企业级 RAG 系统提供了重要的安全范式,避免了完全自动化带来的幻觉风险。
  • 轻量化私有 AI 基础设施:对于无需复杂模型微调但需要深度数据洞察的场景,基于现有基础设施的轻量级网关方案比全栈大模型部署更具经济性和安全性优势。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Agent Agent RAG 检索增强生成