Unify business data, knowledge, memory, context, and provenance.汇聚业务数据、知识、记忆、上下文与来源依据。
Goku Core v1.9.128 · Self-hostedGoku Core v1.9.128 · 私有化部署
From data to decisions. From action to evolution.让数据成为决策依据，
让实践推动自主进化。
Goku does not stop at processing and displaying data. It turns trusted evidence into decisions and controlled actions, learns from what actually happened, and continuously improves the next decision.
Goku 不止处理和展示数据，而是把可信数据转化为决策依据与受控行动；再通过真实实践的结果反馈，持续优化下一次判断，形成能够自主学习、受治理地自主演进的体系。
4product layers产品分层
7+delivery channels触达渠道
3UI languages界面语言
100%self-hostable可私有化
MyPage · COO我的首页 · COOagents monitoring智能体持续监控
MORNING BRIEF · DATA AS OF 06:00今日简报 · 数据截至 06:00
Three signals need your attention.今天有三个信号值得你关注。
Settlement success rate结算成功率-2.8%
Declining for 3 days. 71% of the change comes from two channels.连续 3 天下降，71% 的变化集中在两个渠道。
Evidence查看依据 Create task创建任务Cross-agent investigation多智能体联合调查in progress进行中
Operations, finance, and risk agents are working from one shared task board.运营、财务与风控智能体正在共享任务板上协作。
Observe观察
Reason分析
Act行动
Review复盘
The shift产品转变
From a blank chat box to an operating system for work.从空白对话框，走向企业工作的智能操作系统。
A useful enterprise agent must know what matters, keep state across time, work with others, take controlled action, and show its evidence.
真正有用的企业 Agent，需要知道什么重要、跨时间保持状态、与他人协作、受控行动，并交代清楚证据。
AI assistantAI 助手
- Waits for the right question等待用户提出正确问题
- Answers, then forgets回答完毕，工作就结束
- One model, one conversation一个模型，一段对话
- Governance added later事后再补安全与治理
→
Goku agent runtimeGoku 智能体运行时
- Pre-computes role-specific signals主动生成角色所需信号
- Tracks state, evidence, and outcomes持续跟踪状态、证据与结果
- Delegates and holds working meetings能够委派并召开协作会议
- Permissions, approvals, DLP, audit by design权限、审批、DLP 与审计原生内置
The decision loop决策闭环
Data is the beginning—not the deliverable.数据不是交付终点，而是决策与进化的起点。
Goku connects evidence, judgment, action, outcome feedback, and controlled evolution in one traceable loop. The system gets better because it can compare what it expected with what actually happened.
Goku 将证据、判断、行动、结果反馈和受控演进连成一条可追溯的链路。系统通过比较“原先的判断”与“实践的结果”，不断提升下一轮决策质量。
Explain signals, assess risk, compare options, and give decision grounds.解释信号、评估风险、比较方案并给出决策依据。
Create tasks, call tools, collaborate, and request approval when needed.创建任务、调用工具、协同执行，并在必要时请求审批。
Record outcomes, human feedback, exceptions, cost, and tool performance.记录结果、人员反馈、异常、成本与工具表现。
Turn traces into proposals, experiments, skills, and better future decisions.把轨迹转化为改进提案、实验、技能和更好的下一次决策。
Evidence → decision → action → feedback → evolution → better evidence and decisions证据 → 决策 → 行动 → 反馈 → 演进 → 更好的证据与决策
Built for real operations面向真实运营
A system that learns from work—not only from prompts.不是只从提示词学习，而是从真实工作中持续学习。
Every capability connects to the same runtime record: what evidence supported the decision, which action followed, what outcome returned, and how the next cycle should improve.
所有能力都连接到同一条运行记录：什么证据支持了判断、随后采取了什么行动、结果如何，以及下一轮应该怎样改进。
Role-aware MyPage角色化智能首页
Background agents refresh governed snapshots so leaders open to conclusions, trends, anomalies, evidence, and suggested actions—not an empty prompt.
后台 Agent 刷新可信快照，让管理者打开即见结论、趋势、异常、依据与行动建议。
Snapshot-based · traceable快照式 · 可追溯Stateful work有状态的持续工作
Conversations become tasks with streamed steps, durable state, kanban tracking, schedules, retries, and explicit business transitions.
对话进入任务、实时步骤、看板、调度、重试与持久状态，业务动作拥有明确状态转移。
Chat → task → state → outcome对话 → 任务 → 状态 → 结果Delegation and meetings委派与智能体会议
Agents call specialists or work through a shared meeting board. Human permissions flow through every hop; budgets and depth limits contain work.
Agent 调用专家或在共享任务板上开会；每一跳继承人的权限，并由预算和深度限制控制。
Permission intersection权限取交集Studio for agents智能体构建 Studio
Define agents, DAG workflows, tools, MCP, knowledge, memory, skills, plugins, connectors, and documents in a separate builder plane.
在独立构建平面管理 Agent、DAG 工作流、工具、MCP、知识、记忆、技能、插件和连接器。
Goku Studio v1.9.66Omnichannel全渠道触达
Unify PWA, email, Feishu, WeCom, DingTalk, Teams, LINE and more with identity binding, quiet hours, DLP, signed links, and recovery.
统一 PWA、邮件、飞书、企微、钉钉、Teams、LINE 等渠道，支持身份绑定、静默时段、DLP 与失败恢复。
One notification hub统一通知中枢Autonomous, governed evolution自主学习、受控演进
Goku learns from human feedback, task traces, outcomes, tool failures, and experiments—turning practice into reviewable proposals, reusable skills, and better future decisions. High-risk changes remain human-approved.
Goku 从人员反馈、任务轨迹、实践结果、工具失败和实验中学习，把真实工作沉淀为可审阅提案、可复用技能和更好的下一次决策；高风险变更仍由人审批。
Learn · propose · test · approve · improve学习 · 提案 · 验证 · 审批 · 改进Architecture系统架构
Separate planes. One governed operating model.分层独立，运行治理一体化。
The runtime stays stable while builders evolve agents and the router evolves model policy. Each layer deploys and scales independently.
运行时保持稳定，Studio 演进智能体，Router 演进模型策略；每层都可独立部署和扩缩容。
Goku Core
Chat, MyPage, tasks, approvals, audit, channels, state, collaboration, external API.对话、MyPage、任务、审批、审计、渠道、状态、协作与外部 API。
Goku Studio
Agents, workflows, tools, MCP, knowledge, memory, skills, plugins.Agent、工作流、工具、MCP、知识、记忆、技能与插件。
Goku Router
OpenAI-compatible gateway for workload, policy, cost, quality, fallback, and traces.兼容 OpenAI 的模型网关，按任务、策略、成本与质量路由。
Goku SDK
Python and TypeScript clients, sessions, streamed results, external tools.Python 与 TypeScript 客户端、会话、流式结果与外部工具。
Shared trust boundary:统一信任边界： tenant · identity · policy · evidence · cost · observability租户 · 身份 · 策略 · 证据 · 成本 · 可观察性
Governance that runs运行中的治理
Controls belong in the execution path.控制措施必须进入执行路径。
Goku applies identity, access, data protection, approval, and audit at the point an agent acts.
Goku 在 Agent 行动的当下执行身份、访问、数据保护、审批与审计。
DEFAULT DENY
The human never disappears behind the agent.Agent 背后的人，始终不能消失。
Tenant, department, team, user, and agent context travel with the task. Collaboration uses the intersection of human and agent permissions—never an agent as an access-control shortcut.
租户、部门、团队、用户与 Agent 上下文随任务流转。协作采用人与 Agent 权限的交集，Agent 不能绕过访问控制。
RBAC + Agent PolicyVisibility and action rights可见性与行动权限
DLP / PIIBefore LLM and delivery模型调用与对外发送前
Human ApprovalRisk-tiered and auditable风险分级且可审计
Audit + ReplayCalls, tools, decisions, collaboration调用、工具、决策与协作
“A fallback that stays silent is not resilience. It is hidden failure.”“不发声的降级不是韧性，而是隐藏的故障。”
Goku exposes database, vector store, embeddings, Redis, and LLM routing health—with privileged diagnostics separated from public probes.
Goku 展示数据库、向量库、Embedding、Redis 与模型路由健康度，并分离公开探针和管理员诊断。
Subsystem health子系统健康状态
Database + migrations● healthy
Vector store + dimensions● healthy
LLM Router + breakers● healthy
Where it fits典型场景
For work that crosses systems, people, and time.适合跨系统、跨角色、跨时间的工作。
Start with one governed loop, prove the model, then add agents and channels without replacing the runtime.
从一个受治理的工作闭环开始，再逐步增加 Agent 与渠道，无需重建底座。
Role dashboards, anomaly briefs, evidence, action tracking.角色看板、异常简报、证据与行动跟踪。
Monitoring, alerts, approvals, remediation, review.持续监控、风险预警、审批、整改与复盘。
Omnichannel intake, triage, collaboration, SLA tracking.全渠道受理、分诊、协作与 SLA 跟踪。
Research, RAG, memory, reports, reusable skills.研究、RAG、长期记忆、报告与可复用技能。
Deploy your way按你的方式部署
Keep models flexible. Keep control local.模型保持灵活，控制留在本地。
Run with Docker Compose, Kubernetes and Helm, or an offline bundle. Connect cloud or local models through Goku Router.
支持 Docker Compose、Kubernetes + Helm 与离线包；通过 Goku Router 接入云模型或本地模型。
● ● ●
# One-command evaluation install curl -fsSL https://raw.githubusercontent.com/chenbinsg/goku-core/main/install.sh | bash # Or use the SDK pip install aios-sdk
Put your first governed agent loop into production.让第一个受治理的智能体闭环进入生产。
Architecture review, deployment planning, or a focused proof of concept.从架构评审、部署规划或一个聚焦的 POC 开始。