AgentFlux README

项目的 README,来自 github.com/davemessew/AgentFlux。合并为一个 markdown 文件:README.md。

AgentFlux

AgentFlux

Your data. Your models. Your workflows.

Agentic AI platform for visual workflows and autonomous agents.
Import files, automate tasks and keep results local.

Release Docker License

Download the Docker release · Website · Documentation · Report an issue


AgentFlux overview with workflows, data sources and getting started

What you can do

AgentFlux connects documents, data sources and language models in a visual workflow. Build each step yourself or let an agent create a workflow from a task description. The application runs on your computer or server, with a browser interface.

In AgentFlux
Visual workflows 17 preconfigured steps, eight templates, branching, undo/redo and JSON import/export.
AI agents Set up recurring tasks, reuse a saved plan and answer missing details in text.
Video to workflow Combine screen changes and spoken explanations, then create and run a workflow automatically.
Knowledge and learnings Store documents locally, review observations from runs and reuse instructions as skills.
Your own models Connect Ollama, LM Studio, vLLM and OpenAI-compatible endpoints.
Files and connections Import documents and spreadsheets with named columns; connect folders, HTTP endpoints and approved programs.
Traceable runs Process changed inputs and deliver JSON, CSV and Markdown reports to a local folder.

Create a workflow from a video

Open Agents → Analyse video, select a recording and a model with vision support. AgentFlux analyses overlapping one-minute segments using screen changes, periodic frames and timestamped speech. If an action or integration is unclear, it asks a short question. Answer in text, then save or start the generated workflow. The editor is optional.

Recordings are uploaded to your AgentFlux instance and removed after processing or cancellation. The Docker release includes FFmpeg and a local speech service with the Whisper base model. Images go to your selected vision model. External image and audio processing require separate approval; audio stays local by default.

Videos can be up to 250 MB and 30 minutes. Each segment uses up to 24 selected frames, with explicit call and token budgets for the whole job. Progress shows analysed coverage. Unsupported codecs, exhausted budgets and transcription failures stop the job instead of producing a complete workflow. Brief actions can still be missed.

Generated workflows use existing connections and executable steps. They do not replay arbitrary mouse or keyboard actions in other applications.

Get started with Docker in minutes

You need Docker with Compose. The release includes the application, built interface, FFmpeg and local speech recognition. You do not need to install Node or pnpm locally.

Use the prebuilt image

Download these two files from release v0.3.0 into the same folder:

Open a terminal in that folder:

docker load -i agentflux-0.3.0-linux-amd64.tar.gz
docker compose up -d --wait

Open localhost:4310. Done.

The prebuilt image targets Linux/amd64. On Windows, use Docker Desktop in Linux container mode. You can build from source for other architectures; they were not tested for this release.

Download from the terminal
mkdir agentflux
cd agentflux
curl -fLO https://github.com/davemessew/AgentFlux/releases/download/v0.3.0/agentflux-0.3.0-linux-amd64.tar.gz
curl -fLO https://github.com/davemessew/AgentFlux/releases/download/v0.3.0/compose.yaml
docker load -i agentflux-0.3.0-linux-amd64.tar.gz
docker compose up -d --wait

In Windows PowerShell, use curl.exe instead of curl if needed. Checksums and full instructions are also included in the release.

Prefer to build from source?
git clone https://github.com/davemessew/AgentFlux.git
cd AgentFlux
docker compose up --build -d --wait

Installation without Docker is also documented.

A look inside

The screenshots show the interface included in v0.3.0, using sample data.

Build workflows on the canvas

Connect sources and processing steps, configure each step and follow its execution.

Visual workflow editor with connected steps

Agents
Tasks, models and recurring jobs in one place.
Knowledge base
Search documents and processed content locally.
Manage agents in AgentFlux Local knowledge base with sample documents

Your first workflow

  1. Open Agents → Create agent and choose Clean and export records without a model.
  2. Select Local inbox, choose an output directory and start. Run details show the delivered files.
  3. Connect your model server under Models to extract sourced tasks or write reports.
  4. Enable a schedule to process new or changed inputs automatically. Missing owners and dates remain unresolved.

The release includes a speech model. For agents and video interpretation, connect your own language or vision model server, or an external provider with your credentials.

Model server on your computer Base URL from the container
Ollama http://host.docker.internal:11434
LM Studio / OpenAI-compatible http://host.docker.internal:1234/v1
vLLM http://host.docker.internal:8000/v1

Adjust the port and model ID for your server. The model server must accept connections from the container.

Operations and data

docker compose ps            # Status and healthcheck
docker compose logs -f       # View server output
docker compose stop          # Stop
docker compose up -d --wait  # Start again

Workflows, documents, credentials and run logs stay in the named agentflux-data volume. Reports are stored in its outbox directory by default. The separate agentflux-models volume caches speech assets. Replacing the containers does not delete these volumes. Back up the entire volume including .vault-key; docker compose down -v deletes the volume.

[!IMPORTANT] The supplied configuration binds port 4310 to localhost. AgentFlux has no login and is designed for a single user. Shared or public access requires an access-control layer and suitable network configuration.

Cloud model calls and external HTTP POST steps require approval in Settings and in the workflow. HTTP GET requests can send information in URLs and headers. Optional redaction uses rules and may miss sensitive information. See the installation guide and operations guide.

Documentation and development

Start here Contents
Docker release Prebuilt image, Compose file, source code and SHA-256 checksums
Install the release Download, verify, start and update
Installation and models Docker, native installation and model connections
Workflows and agents Editor, steps, templates, learnings and skills
Operations and configuration Environment variables, import folders, backups and programs
src/          Web interface and visual editor
server/       API, agents, execution and SQLite storage
shared/       Domain types, workflow rules and templates
services/     Local speech recognition
tests/        Contract and media checks
examples/     Sample local program export
scripts/      Docker release checks
docs/         Usage, operations and screenshots

For local development: Node.js 24+, pnpm 11.25.0, pnpm install --frozen-lockfile, then pnpm dev. Run pnpm test for focused checks; pnpm build also checks TypeScript.

Run the Docker check after building:

docker build -t agentflux:0.3.0 .
node scripts/smoke-docker.mjs agentflux:0.3.0

The check uses its own container and volume. It covers the API, built interface, an example run and data persistence when replacing the container.

Licence

Apache License 2.0. Bundled dependencies and fonts retain their respective licences.


Download release · Visit website · Open an issue

AgentFlux (davemessew/AgentFlux)

正在搜索数据库…

这是什么?AgentFlux 是一个可自托管的智能体 AI 平台。你可以用它搭建可视化工作流,把文档、数据源和语言模型连接起来,并在此基础上运行自主智能体。
这是无代码的可视化工作流搭建工具,还是编程框架?它是一个带浏览器界面的无代码可视化工作流搭建工具:你可以把步骤拖到画布上,也可以描述一个任务,让智能体替你组装工作流。
怎么把它跑起来?下载预构建的 Docker 镜像(agentflux-0.2.0-linux-amd64.tar.gz)以及 v0.2.0 发布版中的 compose.yaml,放到同一个文件夹里,运行 docker compose up,然后打开 localhost:4310。
我需要自己准备 AI 模型吗?需要。Docker 镜像不包含模型,你需要连接自己的本地服务,例如 Ollama、LM Studio 或 vLLM,或者使用你自己凭据的、兼容 OpenAI 接口的端点。
Docker 发布版有什么需要注意的地方?预构建的 v0.2.0 镜像只支持 Linux/amd64(Windows 需要使用 Docker Desktop 的 Linux 容器模式);它没有登录功能,只适合单用户使用;视频转工作流功能只有从源码构建时才有。
视频转工作流功能是做什么的?你提供一段屏幕录像,智能体会抽取最多 24 张 JPEG 帧,判断你执行了哪些操作,并提出澄清问题,然后把结果转换成带时间戳、可编辑的工作流。
我可以导入哪些文件类型和数据源?PDF、DOCX、XLSX、CSV、JSON 和纯文本文件,还可以连接文件夹、HTTP 端点以及已获批准的本地程序。
有现成的工作流可以先试用吗?有。它内置 17 个预配置步骤和 8 个模板,其中包括一个“清洗并保存数据”模板,该模板使用合成的示例数据运行,完全不需要任何 AI 模型。
如果让某个步骤调用云端模型,会发生什么?调用云端模型以及发出出站 HTTP POST 请求,都需要你在“设置”和工作流本身中明确批准。另外,可以选配脱敏规则,先去除敏感数据,不过 README 提醒说,脱敏仍可能遗漏一些内容。
大小
1.3 MB
星标
234
时长
22天
提交 · 90 天