agent_flow
一个运行在服务器端的多AI智能体自动编排开发系统
Self-hosted AI agent platform that runs as a single binary. Give AI assistants the ability to read files, run commands, browse the web, and more, all through a chat interface in your browser, with no separate database or frontend setup needed.
AgentFlow is a self-hosted system for running AI agents on a server. You install it as a single binary, open a browser, and interact with AI assistants that can actually do things on your server, such as reading and writing files, running shell commands, searching the web, and controlling a headless browser. It is written in Go, stores everything in SQLite, and embeds its own web interface, so there is no separate database server or frontend build step to manage.
The core idea is that you configure agents, each with a defined role, rules, and set of tools it is allowed to use, and then those agents carry out tasks autonomously through a chat interface. Agents can also call other agents for subtasks. The system includes 28 built-in tools grouped into categories: file operations, command execution, browser automation, web search, memory management, task planning, and workspace document management. You can extend this further through the Model Context Protocol, which lets you plug in additional external tools over a standard interface.
It connects to AI models from OpenAI, Anthropic, and Google Gemini, or any API that follows the OpenAI-compatible format. The workbench streams responses in real time and keeps a buffer of recent events so the conversation can resume cleanly if your browser tab disconnects.
Workspaces isolate projects from each other. Each workspace has its own database and file directory, so agents working on different projects do not share context. There is also a role-based access control system with 34 permission nodes, which is relevant if you are running this for a team rather than just yourself.
Deployment is straightforward: download or build the binary, point it at a config file or set environment variables with an AF_ prefix, and run it. The README provides pre-built binaries for Windows, Linux x86-64, and Linux ARM. You can also build from source with Go 1.25. The default login is admin with password admin123, which the README implies you should change before exposing the service.
Where it fits
- Run AI agents on your own server that can read and write files, execute shell commands, and automate browser tasks without sending data to a third-party service.
- Set up a team AI assistant with role-based permissions so different people have different levels of access to what the agents can do.
- Manage multiple projects in isolated workspaces so AI agents working on one project never see files or context from another.
- Extend agent capabilities by plugging in external tools via the Model Context Protocol without modifying the core system.