登录

bisheng

dataelementApache-2.0活跃维护 有专家点评

开源企业级 Agent 平台，把智能体、工作流、RAG 知识库和模型管理整合在一起，方便企业搭建、部署和运维 AI 应用。

BISHENG is an open-source enterprise Agent platform for building, deploying, and managing AI applications. It combines AI agents, workflows, RAG and knowledge management, desktop agent integration, and vibecoding app hosting with unified model management, evaluation, observability, access control, and enterprise administration.

v3.0.1发布于 2026-11-10 18:44 UTC12,231star +57 / 7 天2,019fork同步于 2026-11-11 05:26 UTC
Agent 平台企业级 AIRAG工作流编排模型管理
对比 GitHub 官网agentaichatbotdshdsh-plugin

专家点评

上手难度：进阶更新于 2026-11-11

私有化部署的企业 Agent 平台，替代自研的 RAG、工作流和权限后台三件套

适用场景

  • 内网办公自动化：把文档审核、知识问答、表单流转做成 Agent 应用，数据不出机房；换成 SaaS 方案会卡在合规和网络隔离上
  • 需要按部门、文件夹、文件三级授权做知识库问答：平台自带 RBAC/ReBAC 和问答溯源，自研这套权限模型的工作量比接平台大得多
  • 工作流里要 Human in the loop 和成环、并行分支：这类编排在纯 LangChain 代码里要自己写状态机，平台里是配置项
  • 想把内部 Vibe App 或小工具挂到统一入口，带密钥扫描和审批发布：省掉自建发布流水线和审计日志

踩坑

  • 机器只有 16G 内存或 4 核时直接撞墙：docker compose 默认拉起 ES、Milvus、Onlyoffice，README 写的是 8 核 32G 起步，且没给 GPU 要求，说明它默认按 CPU 推理或外接模型服务来设计
  • Docker 低于 19.03.9、Compose 低于 1.25.1 时容器起不来，且报错不一定指向版本问题，排查会绕一圈
  • issue 里前端镜像版本报错（创建/查看助手失败）和特定模型（qwen）在 ReAct 模式下的知识库检索问题被反复提出，说明模型适配层和前后端版本对齐不是开箱即稳，升级前要留回滚余地

同类对比

要轻量 RAG 检索层选 RAGFlow、Dify；只要工作流编排和可观测、不要企业权限体系，Dify 或 Langfuse 组合更省事。BISHENG 的价值在权限、审批、审计这一层，不需要这些就别上它。

怎么上手

由 README 提炼，执行前请确认
原文 
git clone https://github.com/dataelement/bisheng.git
cd bisheng/docker
docker compose -f docker-compose.yml -p bisheng up -d
硬件要求按 README 说明

README 要求 CPU >= 8 Core、RAM >= 32 GB，未说明 GPU 与显存要求；默认随平台安装 ES、Milvus、Onlyoffice 等第三方组件。

有 Docker 镜像
下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼

开源企业级一站式 Agent 平台，帮企业私有化落地 AI 应用与办公自动化

开源企业级 Agent 平台，把智能体、工作流、RAG 知识库和模型管理整合在一起，方便企业搭建、部署和运维 AI 应用。

核心能力

  • 任务模式由主 Agent 拆解清单并派子 Agent 并行执行
  • Workflow 支持成环、并行、批量与 Human in the loop
  • 知识空间按空间/文件夹/文件三级权限授权与 AI 问答溯源
  • 应用工场把 Vibe App 经预检、密钥扫描、审批后发布
  • 企业管控含 RBAC/ReBAC、审批、审计与 API Key 权限边界

容易踩的坑

需 Docker 19.03.9+ 与 Docker Compose 1.25.1+，版本不够会起不来；默认会一并拉起 ES、Milvus、Onlyoffice，机器资源要留够；启动后访问 3001 端口，第一个注册用户即系统 admin。

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

5 个版本v3.0.1
全部 release
  1. v3.0.1最新正式版2026-11-10 · 2026-11-10 18:44 UTC

    BISHENG v3.0.1 — patch release

    Bug fixes

    • Workflow input / workspace chat: plain-text Word files failed to parse. Uploading a .docx that contains no images (via the Input node's dialog or form input, or in workspace chat) failed with [Errno 2] No such file or directory: '.../media', and only worked after adding an image to the document. Pandoc creates the media directory only when a document contains images; the image-reference rewrite step listed that directory without checking it exists. Regression introduced in v3.0.0. (7df786081d)

    Upgrade

    Change the image tags in docker/.env from v3.0.0 to v3.0.1 (backend, frontend, sandbox, runtime-manager, app-proxy), then docker compose pull && docker compose up -d. No database migration.

    Full Changelog: https://github.com/dataelement/bisheng/compare/v3.0.0...v3.0.1

  2. v3.0.0v3.0.0：Vibe App Hosting & Integration with DeepSeek Harness2026-11-04 · 2026-11-04 02:53 UTC
    查看更新说明

    🚀 BISHENG 3.0 Release：Vibe App Hosting & Integration with DeepSeek Harness

    🏭 1. Vibe App Hosting

    Overview: Developers can host apps they build locally on the BISHENG platform. After an app is approved and goes live, users with access can open and use it on the platform. This feature is off by default. To turn it on, the ops team enables it in config.yaml and deploys the app runtime components.

    1.1 💻 Local development and deployment: Developers install the bisheng command-line tool and sign in with a service account key. They can then develop and debug apps locally, declare the runtime, resource tier, and the platform models and knowledge bases the app uses in the app manifest, and submit the app to the platform.

    1.2 ✅ App approval and preview: Approvers see app publishing requests in the Approval Center. They can browse the source file tree and preview the app, then approve or reject the request.

    1.3 🛠️ Hosted app management: App owners open a hosted app under Build → Apps. Owners can view the current resource tier, set who can see the app, enter third-party keys, view and edit app data online, and bring the app online, take it offline, or delete it.

    1.4 🏪 Using apps: Once an app is live, employees with access can find it in the app square. They click Enter app to open it in a new tab and use it right away, with no second sign-in.

    1.5 🔌 External interfaces

    • MCP Server: Capabilities already in the open API, including knowledge base management and retrieval, organization and identity lookup, and hosted app status and data, are now also available over the MCP protocol. Any MCP-compatible agent can call these tools.
    • OpenAI-compatible model API: Hosted apps can call the models configured on the platform.
    • bisheng-sdk: Hosted apps can use the SDK to access platform identity, knowledge base retrieval, and attachment storage.

    🖥️ 2. Workspace Redesign

    2.1 🧭 Global sidebar: The desktop workspace now uses a global sidebar on the left. From top to bottom it shows New chat, Knowledge Space, Subscription, Apps, More, the chat list, and user info.

    2.2 🏬 Discover: Discover is added under More in the sidebar. Users can browse knowledge spaces, channels, apps, and skills on one page, organized by tabs.

    2.3 🧩 Tools & Skills center: Tools & Skills is added under More in the sidebar. Users can browse tools and skills by category, and upload their own skills under My skills. Admins can grant built-in tools and enterprise skills to specific people or departments under Build → Tools & Skills.

    2.4 🎙️ Streaming voice input: After a user clicks the voice button in the input box, the transcript appears as they speak. Real-time speech recognition models from Tongyi (Alibaba), Volcengine, Baidu, and Tencent are supported.

    🧠 3. Deep Research and Tools

    3.1 📈 Deep Research quality improvements: The PPT skill adds 5 built-in templates, and generated decks can be previewed online. A new web report skill improves HTML output. Excel file handling is improved. History is compressed and long tool results are offloaded to files automatically, based on the context window of the selected model. Thinking models no longer redo their reasoning as often across multiple calls. Long tasks no longer freeze the browser.

    3.2 📦 Container code execution sandbox: The code executor adds a container sandbox mode. Code runs in a separate container, isolated per session. New installations use the sandbox by default. Upgraded environments keep their current mode; admins can switch by selecting Run in sandbox for the code executor under Build → Tools.

    3.3 🌐 DeepSeek web search and web fetch tool: The web search to

  3. v2.6.0-fix22026-08-11 · 2026-08-11 09:18 UTC
    查看更新说明

    Full Changelog: https://github.com/dataelement/bisheng/compare/v2.6.0-fix...v2.6.0-fix2

    What's Changed

    Full Changelog: https://github.com/dataelement/bisheng/compare/v2.6.0-fix...v2.6.0-fix2

  4. v2.6.0-fix2026-07-30 · 2026-07-30 11:39 UTC
    查看更新说明

    What's Changed

    New Contributors

    Full Changelog: https://github.com/dataelement/bisheng/compare/v2.6.0...v2.6.0-fix

  5. v2.6.0v2.6.0 — Linsight Task Mode & Enterprise Permission Upgrade2026-07-18 · 2026-07-18 16:03 UTC
    查看更新说明

    This release consolidates everything from v2.5 and v2.6.

    🧠 1. Linsight Upgraded to Task Mode Linsight has been re-architected and upgraded, adding task planning, sub-agent delegation, a virtual file system, human-in-the-loop (HITL) collaboration, and skill support — for more stable planning, more accurate tool calls, long processes that are less likely to go off track, faster execution, and better results.

    1.1 📋 Plans Before It Acts: Task mode breaks your request into a visible to-do list and works it top to bottom, so you always see which step it's on and what's next.

    1.2 🤝 Delegates the Deep Digging: Heavier investigation is handed to research sub-agents that work in an isolated context and fold a distilled summary back into the main task.

    1.3 💬 Human-in-the-Loop (HITL): When your intent is ambiguous, the task pauses to ask a clarifying question — with interactive options — and resumes automatically once you answer.

    1.4 🧩 Skills, On Your Terms: Administrators can create, upload (.md / .zip / .skill), or import skills from GitHub, and enable, disable, or delete them in the console.

    1.5 ⏹️ Stop & Resume Across Turns: Halt a running task at any time and continue the same conversation later — subsequent turns can see and build on the files produced earlier.

    🔐 2. Enterprise Permission System Upgraded

    The permission system is upgraded to a combined RBAC + ReBAC model, letting large organizations model their real structure and share with precision. Full feature details >> https://dataelem.feishu.cn/wiki/BWSMwdc2Ciwif4kfe5HcyGHgn3b#share-QWTRdhvMiofvSex2h81c6IR6nBg

    2.1 🌳 Model Your Org as It Really Is: Build a multi-level department tree in the admin console — create, rename, move under a new parent, and delete — with department administrators managing their own department's members and roles, and user groups to gather people from across departments.

    2.2 🎚️ Share at Four Relationship Levels: Whether it's a knowledge space, document library, folder and file, workflow or assistant app, tool, dashboard, or channel, permission management lets you grant Owner / Can Manage / Can Edit / Can View to an individual, a department (optionally its whole subtree), or a user group.

    2.3 🛡️ Permission Guardrails: A non-super-admin can never grant above the level they hold, every resource always keeps at least one owner, and anyone who gains or loses access is notified in-app. After upgrading from an earlier version, existing role-, group-, and space-member permissions migrate into the new model in one operation.

    2.4 🎫 Roles, Menus & Quotas: Policy roles control which feature menus a user sees after login, with three per-role quotas — knowledge-base upload capacity (GB), channels that can be created, and knowledge-space subscriptions.

    ⚡ 3. A Smarter Workstation with Inline Citations In everyday chat, the model now judges when to call a tool (including knowledge-base retrieval) and uses citation badges to mark the source of every grounded answer.

    3.1 🤖 Agentic Tool Calling: With one or more tools or knowledge bases enabled, the model decides whether and when to call them and can chain several calls within a single answer. Administrators can customize which tools are available in daily mode from the admin console.

    3.2 🔎 Inline Source Citations: Across daily mode, workflows, and assistants, answers carry numbered citation badges — blue for knowledge sources, purple for web search. Click one to preview the source (PDFs box-highlight the exact cited passage), or open the References panel under an answer to see every source it drew on.

    3.3 📤 Export & Reuse Conversations: Pick the answers you want and export them as Word (.docx), PDF, Markd

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译
    在 GitHub 打开

    源自中国匠心，希望我们能像 [Deepseek]、[黑神话：悟空] 团队一样，给世界带来更多美好。

    BISHENG是一款 开源的企业级一站式 Agent 平台，专注办公效率与企业 AI 应用落地，已有大量行业头部组织及世界 500 强企业在使用。

    重要更新：BISHENG 客户端重磅发布

    BISHENG 桌面端基于 DeepSeek Harness（DSH）开发，与 BISHENG 平台协同，为企业提供一整套开源、可私有化部署、自主可控的企业级 Agent 平台方案。

    简单来说，你可以把它理解为 WorkBuddy／豆包工作的开源替代方案，让企业把 AI 能力真正掌握在自己手中。

    BISHENG 网页端、桌面端与手机端展示

    下载 BISHENG 客户端

    若需要对客户端进行改造（比如更换 LOGO），可基于客户端开源代码改造。

    BISHENG 团队于 DSH 发布当晚第一时间开源了 DSH 的桌面客户端（dshdesktop.com）。BISHENG 客户端基于 DSHDesktop 开发。截至目前，DSHDesktop 已获得 90 万+下载、17.4K GitHub ⭐️。个人用户可以使用 DSHDesktop。

    DSHDesktop 官网

    特点

    1. 🖥️ BISHENG 客户端：连接本地工作与企业云端

      • Chat 与 Work 两种工作方式：Chat 支持知识问答与云端多步任务；Work 在获准的本地目录中整理材料，生成可编辑的 Word、Excel、PPT，并根据反馈持续修改。
      • 企业资源按身份开放：使用企业账号调用获准的模型、API/MCP 工具、技能与知识空间，也可从应用中心打开已发布的助手和工作流；知识检索继续检查空间与文件权限。
      • 模型与积分集中管理：管理员统一配置模型授权、月度积分和调用记录；Chat 与 Work 的企业模型调用共用当前用户的积分账本，用户可查看余额与用量。
      • 企业扩展统一分发：管理员集中导入插件和工作台，员工按企业配置与授权获取使用。

    2. 🛠️ 应用工场：把 Vibe App 发布给团队

      • 把业务需求做成应用：将报名表、项目台账、部门看板等需求做成专门的页面，让同事从应用中心打开使用。
      • 从本地开发到团队使用：提交本地或由 Coding Agent 辅助开发的应用，经预检、密钥扫描、构建和发布审批后，再设置使用范围，发布到应用中心。
      • 复用平台能力：应用可按需复用平台登录与访问者身份，接入模型和知识资源，减少重复建设；具体能力与配置要求以对应版本说明为准。

      BISHENG 应用工场的 Vibe App 发布管理界面

    3. 🚀 任务模式：面向多步骤复杂工作的通用 Agent 你给出目标，按需附上文件、知识库、工具和技能，它会拆解任务、逐步执行，最后交付报告、PPT、Excel 或 Word。

      • 🧠 任务规划 + 子 Agent 并行：主 Agent 动态生成任务清单、跟踪进度，把独立调研交给子 Agent 并行完成。
      • 🙋 人机交互（Human in the loop）：遇到只有你知道的关键信息，任务暂停等你回答。
      • 🧰 技能（Skill）即插即用：内置 PPT、Excel、Word 三个技能包，开箱即用；也可以从本地或 GitHub 导入团队自己的技能。
      • 📎 结论可溯源：报告带引用角标，点击后打开原文，并定位到被引用的段落。

    4. 📚 知识空间：企业版团队知识库

      • 🗂️ 有序组织：支持多级文件夹、标签、卡片与列表视图，支持关键词和标签联合检索。
      • 👥 团队共建：可以按权限模板（所有者/管理者/编辑者/查看者）邀请成员，也可以发布到知识广场供他人加入。超级管理员可以为部门一键开通「部门知识空间」。
      • 🔐 权限细到单个文件：空间、文件夹、文件三级共用一套权限模板，可以一次授权给个人、部门（含子部门）或用户组。文件夹和文件默认继承上级权限，也可以单独设置。
      • 💬 AI 问答：可以针对单篇文档提问，也可以基于整个空间提问；回答标注引用来源，点开即可查看原文及溯源定位。

    5. 📡 频道订阅：外部资讯每天自动送达，有价值的一键存成知识

      • 📰 主题频道：把公众号和网站聚合成主题频道，每天自动更新，可以再分子频道。
      • 🎯 精准过滤：自定义规则过滤标题和正文，只留下相关资讯。

    6. 🧩 独具特色的 BISHENG workflow

      • 独立、完备的应用编排框架：不区分 chatflow 和 workflow，多轮对话和自动化处理可以混在同一个流程里。
      • Human in the loop：流程执行到中途可以等人输入、确认或修改。
      • 成环、并行、批量运行自由组合：画圈即成环（大多数同类产品不支持成环），并列即并行，多选即批量，适合文档审核、固定版式报告生成等复杂场景。

    7. 🏢 企业级特性：企业 AI 落地需要的管控能力

      • 客户端统一管理：管理员集中配置客户端可用的模型、技能、API/MCP 工具与知识空间，按企业配置和资源权限开放给员工；通过企业插件市场分发插件和工作台，并统一管理月度积分、调用用量与审计记录。

      • 部门和用户管理：维护组织架构、用户和配额。

      • 功能与资源权限分开管理：RBAC 通过角色控制功能菜单与操作；ReBAC 控制知识空间、文件夹和单个文件的访问，可按个人、部门（含下级部门）或用户组授权。文件夹和文件默认继承上级权限，也可独立设置。

      • 操作审批：知识空间加入、频道订阅、菜单权限申请等操作可以走审批，审批流程和节点可自定义。

      • 审计：记录会话内容和系统操作日志，可按人、时间、模块追溯。

      • 统计看板（商业版）：预置用户活跃度、应用使用、token 用量、知识库使用等数据看板。

      • API Key 权限边界：管理员为服务账号签发密钥，配置有效期、接口权限，并授予可访问的应用、知识库等资源。接口权限决定能做什么，资源授权决定能对哪些资源操作；获得密钥不等于拥有整个平台权限。

        BISHENG API Key 签发与权限配置界面

      • 信创适配：同时支持 MySQL 和达梦 DM8。

      • 以及 SSO/LDAP、高可用部署方案、安全审查、流量控制、监控统计、漏洞扫描修复。

      BISHENG 企业管理后台的客户端积分与用量页面

    8. 🔍 高精度文档解析：5年海量数据沉淀，高精度文档解析模型支持免费私有化部署使用，包括高精度印刷体、手写体与生僻字识别模型、表格识别模型、版式分析模型、印章模型...

    快速安装

    安装BISHENG前请先确保满足以下条件：

    • CPU >= 8 Core
    • RAM >= 32 GB
    • Docker 19.03.9+
    • Docker Compose 1.25.1+

    除了BISHENG前后端，我们默认还会安装第三方组件ES、Milvus、Onlyoffice

    下载BISHENG代码

    # 如果系统中有git命令，可以直接下载BISHENG代码
git clone https://github.com/dataelement/bisheng.git
# 进入安装目录
cd bisheng/docker

# 如果系统没有没有git命令，可以下载BISHENG代码zip包
wget https://github.com/dataelement/bisheng/archive/refs/heads/main.zip
# 解压并进入安装目录
unzip main.zip && cd bisheng-main/docker

    启动BISHENG

    # 进入bisheng/docker或bisheng-main/docker目录，执行
docker compose -f docker-compose.yml -p bisheng up -d

    启动后，在浏览器中访问 http://IP:3001 ，出现登录页，进行用户注册。默认第一个注册的用户会成为系统admin。

    其他安装部署问题参考：私有化部署

    感谢

    感谢我们的贡献者：

    社区与支持

    欢迎加入我们的交流群

    服务企业

    BISHENG 已服务众多行业头部企业，以下为部分客户。

    中国气象局、汇丰银行、博世、屈臣氏、中国移动、中粮集团

    首钢集团、中国宝武、中金公司、中国出口信用保险公司、中国石油、中国工商银行

    重庆银行、中国电建、华泰证券、华夏银行、辽宁金控、中国民生银行

    龙华资本、苏交科、中信银行、中英人寿、北汽集团、中债资信

    企业咨询

    如需了解 BISHENG 企业部署、桌面客户端接入、产品演示或商务合作，欢迎联系我们。

    填写飞书表单，联系 BISHENG 团队

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    langfuse
    langfuse·TypeScript

    开源 LLM 应用可观测与评估平台，能追踪调用链路、跑评测实验，帮团队定位问题并持续优化提示词和模型效果。

    LLM 可观测+4LLM 可观测进阶
    37278+1544.2kv4.78.02026-11-11 03:27 UTC
    signoz
    SigNoz·TypeScript

    基于 OpenTelemetry 的开源可观测性平台，把日志、指标和链路追踪集中到一处，提供 APM、分布式追踪和基础设施监控，帮团队快速定位线上问题。

    可观测性+4LLM 可观测进阶
    32579+352.6kv0.149.02026-11-11 05:14 UTC
    skywalking
    apache·Java

    分布式系统的应用性能监控平台，支持链路追踪、指标采集和日志分析，帮你在微服务架构下快速定位性能瓶颈。

    应用性能监控+4LLM 可观测硬核
    25001+76.6kv11.1.02026-11-11 00:56 UTC
    jaeger
    jaegertracing·Go

    CNCF 出品的分布式追踪平台，采集服务调用链路并做可视化分析，帮你在微服务和 LLM 应用里定位延迟与故障根因。

    分布式追踪+4LLM 可观测进阶
    23395+93.2kv2.23.02026-11-10 17:10 UTC