川序 Chuanxu 川序 Chuanxu

AI Agent Infra with DB · v4.4.15

川序ChuanxuAI Agent 管理平台AI Agent Management Platform

水到而渠成，让智能体运行有序、可信As Water Arrives, the Channel Takes Shape: Orderly and Trusted AI Agent Operations

川Chuan寓意多源汇流、持续运行与开放生态Many sources converging, continuous operation, and an open ecosystem 序Xu代表秩序、编排、权限与治理Order, orchestration, authorization, and governance

Agent 从 POC 进入生产，关键不再只是模型能力，而是状态能否恢复、过程能否审计、权限能否治理。川序是基于数据库的 AI Agent 管理平台，统一管理已注册 Agent 的身份、权限、组织归属、任务、协作、运行状态与审计证据，让企业中的 Agent 可观测、可调度、可运维。平台账号、组织人员与 Agent 责任关系以数据库事实对齐；不同框架中的 Agent 通过 Skill 等受控接口接入，核心状态在异常后可由替代实例恢复。As Agents move from POC to production, the critical questions are no longer only about model capability, but whether state can recover, execution can be audited, and authority can be governed. Chuanxu is a database-native AI Agent management platform. It unifies identity, authorization, organization ownership, tasks, collaboration, runtime state, and audit evidence for registered Agents, making enterprise Agents observable, schedulable, and operable. Platform accounts, organization people, and Agent accountability align as database facts; Agents from different frameworks connect through controlled interfaces such as Skills, and replacement instances can recover core managed state after failure.

3 种数据库3 Databases6 个版本6 EditionsLinux 运行时隔离Linux Runtime IsolationDB4A2ASkill 优先Skill-first零信任Zero Trust
了解产品能力Explore the Platform部署社区版Deploy Community预约演示 / 联系合作Book a Demo / Partner
继续了解川序Explore Chuanxu
从 POC 到生产From POC to Production

Agent 增加了，状态与治理没有跟上Agent Adoption Is Outrunning State and Governance

Demo 可以依赖进程内状态和人工看护；生产系统必须在进程或节点异常后继续工作，解释谁授权了什么，并在真实执行边界阻断越权。企业需要持续回答以下问题。A demo can rely on in-process state and manual supervision. Production must survive process or node failure, explain who authorized what, and block excess authority at the real execution boundary. Enterprises need durable answers to these questions.

01有多少？How many?缺少统一 Agent 台账No unified inventory
02谁负责？Who owns them?身份与归属不清晰Identity and ownership are unclear
03在做什么？What are they doing?运行过程仍是黑盒Execution remains opaque
04如何阻断？How do we stop them?提示词不能代替权限Prompts cannot replace authorization
05如何追溯？How do we review?证据分散且不可复核Evidence is fragmented
产品结果Product Outcomes

可观测、可调度、可运维Observable, Schedulable, Operable

以数据库中的权威状态为基础，把分散 Agent 转化为可查询、可授权、可恢复、可审计的企业资产。Use authoritative database-backed state to turn scattered Agents into enterprise assets that can be queried, authorized, recovered, and audited.

可观测OBSERVE

统一看见See the Estate

查看组织、人员、Agent 归属、状态、任务、频道、关系和访问轨迹。Inspect organizations, people, Agent ownership, state, tasks, channels, relationships, and access trails.

可调度ORCHESTRATE

持续调度Coordinate Work

用任务计划、协作关卡、Loop 和 Graph 组织长周期、多 Agent 工作。Organize long-running, multi-Agent work with task plans, collaboration gates, Loops, and Graphs.

可运维OPERATE

安全运维Operate Safely

统一身份、权限、审批、审计、应急控制，以及异常后的状态恢复。Unify identity, authorization, approval, audit, emergency control, and state recovery after failure.

统一管理边界Unified Management Boundary

多个 Agent 接入一个受管平台Many Agents, One Governed Platform

Agent 可以运行在不同框架、主机或业务环境中，通过 Skill、MCP、HTTP 或 CLI 接入川序；进入平台后，其身份、权限、协作和执行才进入统一管理范围。Agents may run in different frameworks, hosts, or business environments. They connect through Skill, MCP, HTTP, or CLI; once inside Chuanxu, identity, authorization, collaboration, and execution enter one managed boundary.

v4.4.15 Production Baseline

隔离 Agent 运行时，减少 Agent 间数据搬运Isolate Agent Runtimes and Reduce Agent-to-Agent Data Movement

v4.4.15 将平台原生 Agent 的 Linux 执行边界与数据库介导协作纳入同一控制面：主机准入、运行隔离、上下文授权和分支溯源都产生可复核证据。v4.4.15 brings Linux execution boundaries for platform-native Agents and database-mediated collaboration into one control plane, producing reviewable evidence for host admission, runtime isolation, context authorization, and branch provenance.

LINUX RUNTIME ISOLATION

平台原生多 Agent 隔离Platform-native Multi-Agent Isolation

面向 RHEL 9.8（Oracle Linux）及对等或更新系统，首次 root 引导完成后交接给受管账号和 root-owned Host Manager。每个 Agent 使用独立 UID/GID、namespace、只读根文件系统、私有 IPC、默认拒绝网络和资源限制；主机预检、证据或配置漂移不满足策略时失败关闭。On RHEL 9.8 (Oracle Linux) and equivalent or newer systems, initial root bootstrap hands control to a managed account and root-owned Host Manager. Each Agent receives a distinct UID/GID, namespaces, read-only root filesystem, private IPC, default-deny networking, and resource limits. Admission fails closed when host preflight, evidence, or drift posture does not satisfy policy.

DB4A2A

数据库介导的引用协作Database-mediated Reference Collaboration

任务分发时传递受治理的上下文引用，而不是在 Agent 之间复制大体量数据；接收方只读取当前授权范围，并可建立带来源、父分支和完整流向的工作分支。DB4A2A 是平台内数据库介导协作，不等同于默认关闭、需单独验证边界的跨组织 A2A 协议。Task delegation passes governed context references instead of copying large payloads between Agents. A recipient reads only its current authorized scope and may create a work branch with source, parent, and full-flow provenance. DB4A2A is database-mediated collaboration inside the platform; it is distinct from cross-organization A2A protocol support, which remains disabled by default and requires independent boundary validation.

真实产品Product Evidence

不是概念图，而是可部署的管理界面A Deployable Management Surface, Not a Concept

川序 v4.4.15 企业管理大屏Chuanxu executive wallboard
22 个主要管理视图primary management views覆盖管理大屏、组织、知识、身份、任务、协作、运行、治理和按需功能配置。Wallboard, organization, Knowledge, identity, work, collaboration, runtime, governance, and scoped capability configuration.
3 × 2 交付矩阵delivery matrix三种数据库，各有社区版和企业版。Three databases, each with Community and Enterprise.
Skill 优先Skill-firstOpenClaw、Hermes Agent 等可通过 Skill 接入。OpenClaw, Hermes Agent, and others connect through Skills.
查看完整产品能力View all capabilities
0默认零信任ZERO TRUST BY DEFAULT
安全边界Security Boundary

Agent 运行安全不能依靠提示词Agent Security Cannot Depend on Prompts

提示词、API 规范、模型策略、频道成员关系和前端菜单只能表达意图，不能授予权限。川序在服务端按当前身份、归属、角色、显式拒绝、安全域、用途、密级和有效期重新决策。Prompts, API contracts, model policies, channel membership, and visible menus express intent but never grant authority. Chuanxu re-evaluates identity, ownership, roles, explicit deny, security domain, purpose, classification, and validity server-side.

  • 身份、授权或凭证缺失、失效、过期时失败关闭Fail closed when identity, grants, or credentials are missing, stale, or expired
  • 高风险访问支持多人审批、职责分离和即时撤销High-risk access supports multi-party approval, separation of duties, and immediate revocation
  • 同一 Agent 可按安全域、频道或任务隔离实例与短期凭证Isolate instances and short-lived credentials by security domain, channel, or task
了解零信任设计Explore Zero Trust
成熟度边界Maturity Boundary

稳定生产与实验能力明确分离Stable Production and Experiments Stay Explicitly Separate

生产配置PRODUCTION PROFILE

全新部署基线Fresh Deployment Baseline

v4.4.15 在身份、组织、数据库原生 SDD、受保护管理频道、版本化记忆、LLM 生命周期治理和 Graph Runtime 核心之上，提供组织关联知识范围、可选模型网关、Token 与成本事实、配额与财务证据以及登录后只读管理大屏。有效能力由版本包、数据库能力矩阵、运行配置和主体授权共同决定。The v4.4.15 fresh baseline combines organization-aware Knowledge, an optional model gateway, Token/cost and quota/financial evidence, and an authenticated read-only wallboard with identity, organization, database-native SDD, protected administration Channels, versioned memory, LLM lifecycle governance, and Graph Runtime core. Effective capability is the intersection of the package, database capability matrix, runtime configuration, and subject authorization.

图能力矩阵GRAPH CAPABILITY MATRIX

受控与禁用能力Controlled and Disabled Capabilities

Graph Runtime 核心与授权检查为 Production；Manifest Draft Import、SLO 只读和 Checkpoint Fork 为 CONTROLLED；Graph Replay、Dynamic Graph Migration、Framework Adapter Execution、A2A 与 OTLP 为 DISABLED。关闭能力不产生路由、Worker、调度或后台活动。Graph Runtime core and authorized inspection are Production capabilities; Manifest Draft Import, read-only SLO views, and Checkpoint Fork are CONTROLLED; Graph Replay, Dynamic Graph migration, framework-adapter execution, A2A, and OTLP are DISABLED. Disabled capabilities create no routes, workers, schedules, or background activity.

PostgreSQL Ecosystem

更多 PostgreSQL 生态数据库选择More Database Choices in the PostgreSQL Ecosystem

基于 PostgreSQL 生态开发、具备相同能力的数据库，包括 IvorySQL、OpenTenBase 和 Halo DB（羲和数据库）。Databases developed within the PostgreSQL ecosystem with equivalent capabilities include IvorySQL, OpenTenBase, and Halo DB.

Memory Ecosystem

兼容更多记忆模块Compatible Memory Modules

平台兼容 PowerContext 与 TencentDB Agent Memory，支持同时适配运行。The platform is compatible with PowerContext and TencentDB Agent Memory and supports running both integrations together.

下一步Next Step

选择适合你的入口Choose Your Path

企业用户Enterprise Teams

围绕真实 Agent、权限和审计场景安排演示或 4 周以内 POC。Plan a demo or a POC within four weeks around real Agents, access, and audit scenarios.

联系演示Book a Demo

技术用户Technical Evaluators

免费部署社区版，选择 Oracle、PostgreSQL 或崖山数据库验证核心能力。Deploy Community for free on Oracle, PostgreSQL, or YashanDB and validate the core.

开始部署Start Deploying

生态伙伴Ecosystem Partners

开展数据库适配、联合验证、客户引荐和集成交付合作。Collaborate on database adaptation, joint validation, customer introductions, and delivery.

讨论合作Discuss Partnership