🔶 AgentSwarms：可自托管的 Agentic AI/BI 平台，提供沙箱化 Python（ELv2）

AgentSwarms 是一个完整、可自托管、源码可用的 Agentic AI 与商业智能平台，提供 Agent 对话、可视化多 Agent Swarm 画布、带 RAG 的知识库、工具调用、MCP 连接、已连接的数据库与数据仓库、仪表盘、预算管理和完整执行追踪。它专为便捷部署而设计：使用一个 Supabase 项目作为后端，通过一条 Docker 命令即可运行，并支持自带一切——数据保存在你自己的 Supabase 项目中，模型则使用你自己的提供商密钥运行（OpenRouter、OpenAI、Anthropic、Gemini、Bedrock、Azure、OCI、Qwen、Grok、Groq、Ollama、vLLM……）。你也可以选择配置一个实例范围的 OpenRouter 密钥，让实例中的所有用户无需任何初始配置即可开始使用。

Swarm 画布——以图结构设计多 Agent 工作流，并端到端运行。每个节点都是一个步骤（Agent、路由器、条件、循环、审批、工具调用）；检查器可设置其提供商、模型、提示词、工具和知识库。同一张图既可从画布运行，也可通过 API 或按计划运行。

BI 工作区——基于已连接表和数据仓库构建多页面仪表盘，支持 KPI、交叉筛选、定时刷新、PDF 导出以及发布分享链接。

带可视化 BI 的 Agent Chat——使用自然语言提问，即可在回答旁获得基于你自身数据计算的图表。生成该图表的 SQL 会作为来源展示，因此数据结果可被核查，而非仅是声明。

生成文档的 Agent Chat——将对话和你的数据转化为真实、可编辑的 PowerPoint、Word 文档或 Excel 工作簿。Excel 可通过实时公式提取全部行数据，而不是仅提供粘贴的快照。

开发者工作区——在沙箱化服务器内核上运行 Python Notebook，预装真实可用的 langchain、langgraph 和 llama_index。模型和知识库调用由平台代理，因此沙箱中绝不会存在提供商密钥。Notebook 可调用已部署的 Agent，也可自行发布为可调用 API。

本仓库与 agentswarms.fyi

相同的 UI，两种不同的使命：

| 本仓库（源码可用，Elastic License 2.0） | agentswarms.fyi（托管版）
---|---|---
重点 | 在你自己的基础设施上便捷部署完整 Agentic AI 与 BI 平台——Agent、Swarm、RAG、已连接数据、仪表盘、追踪、预算。 | 学习优先：面向 Agentic AI 的实践课堂——引导式课程、跟做实验室、交互式 Notebook、演示文稿和认证——全托管。
运行环境 | 你的 Supabase 项目、你的提供商密钥、你的 Docker 主机（或 Cloudflare Workers）。 | 托管基础设施，包括提供免费层模型的 AI 网关——无需配置。
额外功能 | 对自己数据进行无头控制；除自身预算外没有使用上限。 | 仅托管版提供的功能：现场工程博客、社区画廊、语音 Agent、网站嵌入及免费的独立工具。
最适合 | 希望运行自己拥有的 Agentic AI 平台的团队和技术爱好者。 | 希望无需配置即可学习和实践 Agentic AI 的学习者。

“AgentSwarms”名称和托管服务仍归项目作者所有。

|
---|---
🤖 Agent Chat | 构建一个 Agent、连接工具，并在浏览器中与其对话（位于 Build → Agent Chat），提供完整的请求/响应追踪。开启 Visual BI 后，可在回答旁从已连接表中渲染图表；还可根据你的提示词和对话生成完全可编辑的 PowerPoint、Word 或 Excel——Excel 可通过实时公式提取全部行。参见 Agent Chat 与文档生成。
🐝 Swarm 画布 | 可视化设计多 Agent 工作流（基于 XYFlow），并端到端执行——可从画布、API 或按计划运行。已部署的运行会在执行过程中持续创建检查点，因此一次运行可在重启或部署后存活；人工审批步骤会暂停运行，直至有人作出决定，而不会敷衍批准或直接失败。
📚 知识库 / RAG | 上传文档，对其分块并生成嵌入向量（pgvector），使 Agent 基于你自己的数据进行回答。
🏢 数据源 | 连接 PostgreSQL（Supabase、RDS、Neon……）、MySQL/MariaDB、Oracle（Autonomous DB / ORDS）、Amazon Redshift、Snowflake、Databricks、Google BigQuery、Azure Synapse、Trino / Starburst / Presto 和 Amazon Athena——凭据加密且只读。可通过 Data Catalog SQL 工作台、SQL Agent 查询这些数据源，并为 BI 图表、本体和定时刷新提供数据。参见 数据源与连接器。
🔑 Secrets Manager | 一次性存储凭据（加密、只写），并可在任意位置通过 {{secret:NAME}} 引用——包括数据仓库连接和提供商密钥。超级管理员可通过 IAM 与用户/群组共享密钥。
🗂️ Data Catalog | 通过向导连接数据仓库、S3 兼容存储桶（AWS S3、Google Cloud Storage、Cloudflare R2、MinIO、Spaces、B2）或 Iceberg REST catalog；爬虫会列出全部表和对象，将分区文件夹归组为数据集，通过采样推断 CSV/JSON schema，分析列特征（空值百分比、去重计数、取值范围），估算行数，并标记可能包含 PII 的列。可安排每日/每周增量爬取并接收 schema 漂移通知，使用 AI 生成资产与列文档，以所有者和标签认证或弃用资产，追踪血缘和使用情况（哪些仪表盘、数据准备流和指标使用每张表），定义业务术语表，并直接跳转至 SQL 工作台。
📊 商业智能 | 提供带拖放功能的 BI 工作区：可从本地数据集或已连接的数据源构建图表，使用 AI 分析师生成可视化内容（或完整仪表盘），随后通过公开链接发布，或与 IAM 群组共享。包括企业级能力：每类图表均支持点击交叉筛选和下钻（包括地图、矩形树图、热力图），可钻取至底层行数据，支持区域设置/货币数字格式、带日期预设和固定默认值的仪表盘筛选器、带小计的可展开矩阵（透视表）、支持恢复的版本历史、带邮件报告和“变化内容”洞察摘要的定时刷新、数据警报（应用内 + 邮件）、共享仪表盘的行级安全、使用分析及移动端堆叠布局。可将仪表盘整理到 工作区和文件夹中，按群组进行只读共享，将个人草稿提升至共享工作区，并将模型/仪表盘定义导出到 Git 仓库（GitHub/GitLab）。
🔍 可观测性 | 在完整执行追踪中检查每一次工具调用、Token 和成本；还提供“谁做了什么”的审计记录（模型调用、数据集和数据仓库查询、仪表盘查看、目录爬取），支持配置保留期限，并为管理员提供按用户和 IAM 群组细分的支出分析。
🌐 网页搜索与浏览 | 为 Agent 提供 web_search 和 web_browse 工具。无需密钥即可开箱即用（免费 DuckDuckGo 备用方案——能力有限）；如需真实的工作区级搜索和完整页面读取，可在 Integrations 页面连接 Firecrawl（或设置 FIRECRAWL_API_KEY），也可以为每个 Agent 自带 Brave / SerpAPI / Tavily / ScrapingBee 密钥。每次由模型驱动的抓取都受到 SSRF 防护。参见 网页搜索与浏览。
🔌 BYOK + MCP + A2A | 加密的按用户提供商密钥、MCP 服务器连接、将 Swarm 导出至 LangGraph/CrewAI/OpenAI SDK/Strands，以及 A2A 端点。
🛠️ MCP Builder | 在 Build → MCP Builder 下使用 FastMCP 以 Python 编写 MCP 服务器，将其部署至 Developer workspace 所使用的同一沙箱化内核，并获得真实的 Streamable-HTTP 端点。它默认可缩容至零（或保持预热），会自行注册以便你的 Agent 调用，并可通过支持过期时间、按工具限额和按 IP 限额的哈希 API 密钥公开暴露。密钥会作为环境变量绑定，但绝不会进入容器环境；任何工具名称、描述或 schema 发生变化的重新部署，都会阻止调用，直到你再次批准——这正是 MCP 自身安全指南所要求的防“rug pull”控制。需要启用服务器运行时。
🛂 IAM | 支持超级管理员、群组、邀请/手动用户配置、按用户/群组设置模型允许列表、对知识库和数据表进行只读共享，以及仅邀请模式。
📓 开发者工作区 | 在沙箱化服务器内核上运行 Python Notebook——是真实 CPython，支持可用的 pip install 以及真实的 LangChain、LangGraph 和 LlamaIndex 导入，并非浏览器模拟。内置每个框架的只读可运行示例，以及一个混合 Agentic 技术栈的综合示例（知识库、工具、技能、护栏、MCP）——可 Fork 任意示例进行编辑。内置 agentswarms 辅助程序可调用已连接模型、搜索知识库并运行已保存的 Agent 和 Swarm；一切均受 IAM 规则约束并记录在 Traces 中——沙箱内绝不会存在提供商密钥。Notebook 还可发布为可调用 API，并作为纯 Python 代码版本化至 Git。运营人员可在 Admin → Developer runtime 下启用运行时；参见 运行时指南。
🛡️ 护栏与评估 | 可针对你自己的 Agent 运行提示词注入测试、PII 脱敏和 LLM-as-judge 评分。

一条命令完成设置——在创建 Supabase 项目并将其密钥填入 .env 后（见下文），脚本将处理其余工作（密钥、依赖、迁移和启动技术栈）：

cp .env.example .env      # 填入 Supabase 密钥，然后：
bash scripts/setup.sh                 # Docker 技术栈  →  http://localhost:8080
# bash scripts/setup.sh --dev         # 改为使用本地开发服务器
# bash scripts/setup.sh --docgen      # + 服务端 PowerPoint / Word / Excel 渲染器
# Windows PowerShell:  powershell -ExecutionPolicy Bypass -File scripts\setup.ps1

或者手动完成——无需安装独立后端，因为 Supabase 就是后端（Postgres + Auth + Storage），可作为免费层托管项目运行，而不必自行安装任何组件：

git clone https://github.com/AgentSwarms-fyi/agentswarms.git
cd agentswarms
npm install
cp .env.example .env   # 填入 Supabase + 提供商密钥
# 一次性应用数据库 schema：npx supabase login && npx supabase link && npx supabase db push
npm run dev            # → http://localhost:8080

使用 Docker 自托管（任意支持 Node 的主机——VPS、Fly、Railway、Render、K8s）：

cp .env.example .env   # 填入 Supabase + 密钥，并一次性应用迁移
docker compose up --build
# → http://localhost:8080

首次使用？请参阅**完整安装指南**——其中涵盖 macOS、Linux 和 Windows 上的每一步操作，包括 Supabase 控制台点击步骤，以及针对用户实际会遇到错误的故障排除部分。

文档位于 docs/，每个主题均有一份专门指南：

AgentSwarms is a complete, self-hostable, source-available agentic AI & business-intelligence platform: agent chat, a visual multi-agent swarm canvas, knowledge bases with RAG, tool use, MCP connections, connected databases & warehouses, dashboards, budgets, and full execution traces. It is designed for easy deployment : one Supabase project as the backend, one Docker command to run, and bring-your-own-everything — your data lives in your own Supabase project, and models run against your own provider keys (OpenRouter, OpenAI, Anthropic, Gemini, Bedrock, Azure, OCI, Qwen, Grok, Groq, Ollama, vLLM…). Optionally set one instance-wide OpenRouter key so every user on your instance can start with zero setup.

Swarm canvas — design a multi-agent workflow as a graph and run it end to end. Each node is a step (agent, router, condition, loop, approval, tool call); the inspector sets its provider, model, prompt, tools and knowledge. The same graph runs from the canvas, from the API and on a schedule.

BI Workspace — multi-page dashboards over your connected tables and warehouses, with KPIs, cross-filtering, scheduled refresh, PDF export and publish-and-share links.

Agent Chat, with Visual BI — ask a question in plain language and get a chart computed from your own data beside the answer. The SQL that produced it is shown as the source, so the number is checkable rather than asserted.

Agent Chat, generating documents — turn the conversation and your data into a real, editable PowerPoint, Word document or Excel workbook. The Excel can pull every row with live formulas rather than a pasted snapshot.

Developer workspace — Python notebooks on sandboxed server kernels with real langchain, langgraph and llama_index installed. Model and knowledge-base calls are brokered by the platform, so no provider key ever exists inside the sandbox. Notebooks can call your deployed agents, and can themselves be published as callable APIs.

This repo vs. agentswarms.fyi

Same UI, two different missions:

| This repository (source-available, Elastic License 2.0) | agentswarms.fyi (hosted)
---|---|---
Focus | Easy deployment of the full agentic AI & BI platform on your own infrastructure — agents, swarms, RAG, connected data, dashboards, traces, budgets. | Learning first : a hands-on classroom for agentic AI — guided curriculum, build-along labs, interactive notebooks, presentations, and certification — fully managed.
Runs on | Your Supabase project, your provider keys, your Docker host (or Cloudflare Workers). | Managed infrastructure, including an AI gateway with free-tier models — nothing to configure.
Extras | Headless control of your own data; no usage caps other than your own budgets. | Hosted-only surfaces: field-engineering blog, community galleries, voice agents, website embeds, and free standalone tools.
Best for | Teams and tinkerers who want to run an agentic AI platform they own. | Learners who want to study and practice agentic AI without setting anything up.

The "AgentSwarms" name and the hosted service remain with the project author.

|
---|---
🤖 Agent Chat | Build an agent, wire up tools, and chat with it in-browser (under Build → Agent Chat), with full request/response traces. Flip on Visual BI to render a chart from your connected tables next to the answer, and generate a fully-editable PowerPoint, Word or Excel from your prompt + the conversation — the Excel can pull all rows with live formulas. See Agent Chat& document generation.
🐝 Swarm canvas | Design multi-agent workflows visually (built on XYFlow) and execute them end-to-end — from the canvas, from the API, or on a schedule. Deployed runs checkpoint as they go , so a run survives a restart or deploy, and a human-approval step parks the run until someone decides rather than rubber-stamping it or failing.
📚 Knowledge Base / RAG | Upload documents, chunk and embed them (pgvector), and ground agents in your own data.
🏢 Data Sources | Connect PostgreSQL (Supabase, RDS, Neon…), MySQL/MariaDB , Oracle (Autonomous DB / ORDS), Amazon Redshift, Snowflake, Databricks, Google BigQuery, Azure Synapse, Trino / Starburst / Presto , and Amazon Athena — encrypted credentials, read-only. Query them from the Data Catalog SQL workbench, from SQL agents, and feed BI charts, ontologies and scheduled refreshes. See Data sources& connectors.
🔑 Secrets Manager | Store credentials once (encrypted, write-only) and reference them anywhere as {{secret:NAME}} — warehouse connections, provider keys. Superadmins share secrets with users/groups via IAM.
🗂️ Data Catalog | Connect warehouses, S3-compatible buckets (AWS S3, Google Cloud Storage, Cloudflare R2, MinIO, Spaces, B2) or an Iceberg REST catalog through a wizard; the crawler lists every table and object, groups partitioned folders into datasets, infers CSV/JSON schemas by sampling, profiles columns (null %, distinct counts, ranges), estimates row counts, and flags likely-PII columns. Schedule daily/weekly incremental crawls with schema-drift notifications, generate asset + column documentation with AI, certify or deprecate assets with owners and tags, trace lineage and usage (which dashboards, prep flows and metrics consume each table), define a business glossary, and jump straight into the SQL workbench.
📊 Business Intelligence | A BI Workspace with drag-and-drop dashboards: build charts from local datasets or connected data sources, generate visuals (or whole dashboards) with the AI analyst, then publish with a public link or share with IAM groups. Enterprise depth included: click-to-cross-filter and drill-down on every chart type (incl. maps, treemaps, heatmaps), drill-through to underlying rows, locale/currency number formatting, dashboard filters with date presets and pinned defaults, expandable matrix (pivot) with subtotals, version history with restore, scheduled refreshes with email reports and "what changed" insight digests, data alerts (in-app + email), row-level security on shared dashboards, usage analytics, and a mobile-stacked layout. Organize dashboards into workspaces & folders with read-only group sharing, promote a personal draft into a shared workspace, and export model/dashboard definitions to a Git repo (GitHub/GitLab).
🔍 Observability | Inspect every tool call, token, and cost in a full execution trace — plus an audit trail of who did what (model calls, dataset & warehouse queries, dashboard views, catalog crawls) with a configurable retention window, and admin-only spend analytics broken down by user and IAM group.
🌐 Web search & browsing | Give agents the web_search and web_browse tools. Works out of the box with no key (free DuckDuckGo fallback — limited); for real workspace-wide search + full-page reads, connect Firecrawl on the Integrations page (or set FIRECRAWL_API_KEY), or bring your own Brave / SerpAPI / Tavily / ScrapingBee key per agent. Every model-driven fetch is SSRF-guarded. See Web search& browsing.
🔌 BYOK + MCP + A2A | Encrypted per-user provider keys, MCP server connections, swarm export to LangGraph/CrewAI/OpenAI SDK/Strands, and an A2A endpoint.
🛠️ MCP Builder | Write an MCP server in Python with FastMCP under Build → MCP Builder , deploy it to the same sandboxed kernel the Developer workspace uses, and get a real Streamable-HTTP endpoint. It scales to zero by default (or stays warm), registers itself so your own agents can call it, and can be exposed publicly with hashed API keys that support expiry, per-tool and per-IP limits. Secrets bind as environment variables without ever entering the container environment, and a redeploy that changes any tool name, description or schema blocks calls until you re-approve — the anti "rug pull" control MCP's own security guidance asks for. Needs the server runtime enabled.
🛂 IAM | Superadmins, groups, invite/manual user provisioning, per-user/group model allow-lists, read-only sharing of KBs and data tables, invite-only mode.
📓 Developer workspace | Python notebooks on sandboxed server kernels — real CPython with working pip install and genuine LangChain , LangGraph and LlamaIndex imports, not a browser emulation. Ships with read-only, runnable samples for each plus a mixed agentic-stack capstone (knowledge base, tools, skills, guardrails, MCP) — fork any of them to edit. The built-in agentswarms helper calls your connected models, searches your knowledge base and runs your saved agents and swarms , all governed by IAM rules and logged in Traces — no provider key ever exists inside the sandbox. A notebook can also be published as a callable API , and versioned to Git as plain Python. Operators enable the runtime under Admin → Developer runtime ; see the runtime guide.
🛡️ Guardrails & evals | Prompt-injection tests, PII redaction, and LLM-as-judge scoring you can run against your own agents.

One-command setup — after you've created a Supabase project and put its keys in .env (see below), a script handles the rest (secrets, deps, migrations, and bringing up the stack):

cp .env.example .env      # fill in your Supabase keys, then:
bash scripts/setup.sh                 # Docker stack  →  http://localhost:8080
# bash scripts/setup.sh --dev         # local dev server instead
# bash scripts/setup.sh --docgen      # + server-side PowerPoint / Word / Excel renderer
# Windows PowerShell:  powershell -ExecutionPolicy Bypass -File scripts\setup.ps1

Or do it by hand — there is no separate backend to install, since Supabase is the backend (Postgres + Auth + Storage), run as a free-tier hosted project rather than installing anything yourself:

git clone https://github.com/AgentSwarms-fyi/agentswarms.git
cd agentswarms
npm install
cp .env.example .env   # fill in your Supabase + provider keys
# apply the database schema once: npx supabase login && npx supabase link && npx supabase db push
npm run dev            # → http://localhost:8080

Self-host with Docker (any Node-capable host — VPS, Fly, Railway, Render, K8s):

cp .env.example .env   # fill in Supabase + keys, apply migrations once
docker compose up --build
# → http://localhost:8080

First time? Follow the full installation guide — it covers every step on macOS, Linux, and Windows, including the Supabase dashboard clicks and a troubleshooting section for the errors people actually hit.

The docs live in docs/, one focused guide per topic:

Contributions are welcome — see CONTRIBUTING.md for the workflow, and please read the Code of Conduct first.

Found a vulnerability? Please see SECURITY.md for how to report it responsibly instead of opening a public issue.

AgentSwarms is source-available under the Elastic License 2.0 (ELv2). In plain terms: you may freely use, self-host, modify, and redistribute it — but you may not offer it to third parties as a hosted or managed service , and you may not remove the licensing/copyright notices. A separate commercial license is available from the author for use cases ELv2 doesn't permit (including running it as a SaaS) — reach out if that's you.

The "AgentSwarms" name, logo, and the hosted service are trademarks of the project author and are not licensed for your use; ELv2 covers the code, not the brand.

Every direct dependency uses a permissive license (MIT / Apache-2.0 / ISC / BSD), compatible with redistribution under ELv2 — the full audit and credits for the open-source projects AgentSwarms builds on live in ACKNOWLEDGEMENTS.md.

Built with TanStack Start and Supabase — an agentic AI & BI platform you own.