Home

OpenAgentFleet

Self-hosted computer-use agents. Every agent gets its own disposable Linux desktop — watch it work live in your browser, click into the screen to take over at any moment, and get a phone alert when an agent needs a human.

This wiki is the how-to layer. It gets you running, answers the questions everyone asks, and points into the repository's docs/ for reference material — architecture, security model, API — which lives with the code so it can't drift from it.

Start here

I want to…	Go to
Run it in 10 minutes	Quickstart
Talk to the whole fleet and use / commands	Quickstart (§4)
Give Oaf a folder on my PC or phone and work in sessions	Quickstart (§4c)
Point it at my own models (Ollama, OpenAI, Anthropic…)	Connecting-Models
Arrange agents in an org chart and hand out tickets	Org-Chart-Tickets-and-External-Agents
Put Claude Code, Codex, Hermes or OpenClaw beside the desktops	Org-Chart-Tickets-and-External-Agents (§3)
Script it from Python / the terminal	Python-SDK-and-fleetctl
Boot a sandbox as a real virtual machine	The-QEMU-VM-Tier
Restrict what an agent can reach on the network	Egress-Policies
Fix something that isn't working	Troubleshooting
The questions everyone asks	FAQ

What it is, in one diagram

  Flutter Companion App               Admin Console (React 19)
             \                                  /
        ┌─────▼────────────────────────────────▼─────┐
        │             Orchestrator (Go)              │
        │   Fleet Manager · Agent Loop · Vault       │
        │   Refinement Engine · Multi-Model Gateway  │
        └─────┬────────────────────────────────┬─────┘
              │ Docker API                     │ HTTP / internal net
        ┌─────▼───────┐                 ┌──────▼───────────────┐
        │ Sandbox     │       ...       │ Sandbox              │
        │ XFCE + Xvfb │                 │  agentd (AT-SPI/GUI) │
        │ noVNC proxy │                 │  persistent Py REPL  │
        └─────────────┘                 └──────────────────────┘

Each sandbox is a hardened container by default, or a real VM with "driver": "qemu". The agent perceives through screenshots and the accessibility tree, and acts through a real mouse and keyboard.

Reference documentation (lives with the code)

• Architecture — components, agent loop, data model, networking
• Security model — what the isolation is and is not worth, stated plainly
• HTTP API — every route, with roles
• Roadmap — built vs. planned
• Agent harness — how the loop, prompt and action vocabulary fit together
• Changelog — 1.0.0 and the full pre-release development log

The honesty contract

The README's Design limits, stated plainly section says what each capability deliberately does not do. This wiki follows the same rule: if a page here claims something the code doesn't do, that's the bug we most want filed.