OpenAgentFleet
GitHub

Open source · MIT · runs on your machines

Give the model
a computer.

Every agent gets its own Linux desktop, a mouse and a keyboard. Watch it work live in your browser. Click into the screen and you are driving. Let go and it carries on.

Works beside Claude CodeCodexHermesOpenClaw

Get started Watch the trailer

Move your mouse over the desktop to take over.

Builder · live af-2f74f6ae-0bd You're driving
stepaction

    The trailer

    Seventy-four seconds. Sound on.

    Desktops that drive themselves, an org chart that hands work down, Claude Code, Codex, Hermes and OpenClaw on the same board, and the phone that pings you when an agent is stuck.

    Made for sharing: the vertical cut (36 s, 9:16) · the full trailer (mp4)

    Perceive

    It looks.

    Each turn the agent gets a screenshot with a numbered badge over everything it could click, and the accessibility tree behind them. It reads the screen the way you do, only with labels.

    Decide

    It answers with one action.

    A thought, a verb, a target. One JSON object under a grammar the model cannot escape. No prose, no tool soup, no “let me try a few things”.

    Act

    It clicks.

    agentd, the only process allowed to touch the mouse, keyboard and framebuffer, carries the action out. Then the loop starts over, screenshot first, for as many windows as the goal takes.

    Fleet

    One becomes a fleet.

    Say what you want in one chat and the bots divide it. Builder ships, publishes to the shared catalog and hands to Checker. Checker opens Builder's machine at http://af-2f74f6ae:8000 and reports back. Every bot on its own computer.

    Operator

    Take over at any moment.

    The desktop streams live. Click into it and you are holding the mouse; let go and the agent carries on. A weird dialog takes you two seconds. Want to teach it the task? Do it once with the recorder on.

    Models

    Your GPU. Your models.
    Nothing leaves the building.

    Point the fleet at Ollama, llama.cpp, LM Studio or vLLM on your own hardware and a run costs $0.00. Cloud models plug in behind the same interface with per-turn cost accounting. Mix them by role: one model for vision, another to reason, a cheap one to summarise.

    • Any OpenAI-compatible endpoint, plus Anthropic, Gemini and native Ollama.
    • Role combinations: vision, reasoning, chat, summarise, refine, each its own model.
    • Calibrated automatically: the fleet learns whether a model's coordinates are pixels or fractions, because they will not tell you.
    Recording · Odoo — Mozilla Firefox
    1. click “Reporting” link
    2. click “Sales” menu item
    3. key ctrl+f
    4. type “August”
    5. click “Export” button
    # SKILL.md — Export the monthly sales report
1. focus    "Odoo — Mozilla Firefox"
2. click    "Reporting"          # role: link
3. click    "Sales"              # role: menu item
4. key      ctrl+f · type "August"
5. click    "Export"             # role: button
6. wait_for "Download complete"

    Teach by doing

    Do it once.
    It becomes a skill.

    Turn the recorder on and do the task yourself in the agent's desktop. Every click, keystroke and the accessible element behind it is captured and compiled into a skill the agent follows next time, by label rather than by pixel, so a moved button does not break it.

    Companion app

    It escalates like an employee,
    not a cron job.

    A CAPTCHA, an MFA prompt, a payment screen: the agent parks the task and pings your phone with the screen frozen at that moment. Answer from the lock screen and it resumes. The same fleet chat, sessions and voice mode are in your pocket, with Oaf.

    • One chat with the whole fleet. Name a bot and it answers first.
    • Sessions bound to a device and a folder, like a coding agent, from the phone or with fleetctl host.
    • Voice mode listens, sends, and reads the answer back.
    Oaf, the OpenAgentFleet mascot: a small orange robot at a control panel
    The companion app's fleet chat in light mode The companion app's fleet chat in dark mode
    needs you2m ago
    Checker is on a payment screen

    Frozen at step 41. Approve, take over, or tell it what to do.

    Org chart and tickets

    Give them a chart.
    Hand out tickets.

    Agents report to each other, and work is tickets with an owner. Ask the fleet for something and every part becomes a ticket that waits on what it needs, with the reason it matters in its brief. Stuck work goes to the agent's manager, then to you.

    • Reviews that decide. A reviewer passes or fails the work, and a fail goes back with the findings.
    • Claude Code, Codex, Hermes, OpenClaw. The agents you already use sit in the same chart, working in a folder on your PC, and take the same tickets.
    • Nothing left ownerless. Every open ticket is checked for a next move, and a verifier reopens what is not really finished.
    • Budgets that stop. A monthly ceiling per agent holds it at 100%.

    Bring your agents

    Claude Code, Codex, Hermes, OpenClaw.
    Same fleet. Same tickets.

    The agents you already use join beside the desktops this fleet provisions. They sit in the org chart, take tickets, review each other's work and report what they cost, and each one runs as itself.

    • It runs where it lives. fleetctl host finds the CLIs on your PC; each agent works in a folder you pick, and asks in your terminal before a run unless you say otherwise.
    • Its work is shared. The files it makes are published to the fleet, so a desktop bot can open and review them.
    • Its own permissions. Edits only, or full rein inside its folder. Claude Code runs without your personal settings.
    • Stoppable and resumable. Stop it from the console; the next run on the ticket picks its session back up.

    Claude, Codex, Hermes Agent and OpenClaw are trademarks of their respective owners. OpenAgentFleet is not affiliated with them.

    Security

    Sealed.

    Every stored credential is sealed with AES-256-GCM under one master key you hold. Agents borrow a session, never a password. Each desktop is a container under cgroup limits with swap off, an optional GPU, and an optional egress policy in its own network namespace.

    Read the threat model: what the isolation is worth, and what it is not.

    The console

    This is the real thing.

    A real XFCE session in a hardened container, streaming into the console, and the org chart, the Work board and the agents you bring, as they are. Fleet, recording studio, engine chain and resolution centre live beside them.

    localhost:8081 · Fleet · Builder
    The OpenAgentFleet console with an agent's desktop streaming live, light modeThe OpenAgentFleet console with an agent's desktop streaming live, dark mode
    The org chart: desktop bots and a Claude Code agent reporting to Builder, light modeThe org chart: desktop bots and a Claude Code agent reporting to Builder, dark mode
    The Work board: tickets by status, with verify tickets and their verdicts, light modeThe Work board: tickets by status, with verify tickets and their verdicts, dark mode
    Adding a Claude Code agent: the PC, its folder, and how much it may do alone, light modeAdding a Claude Code agent: the PC, its folder, and how much it may do alone, dark mode

    Get started

    Ten minutes to a fleet.

    You need Docker with Compose v2 and make. Everything else runs in containers, including the sandbox image it builds for you.

    Run the stack
    $ git clone https://github.com/BryantVanOrden/OpenAgentFleet.git
$ cd OpenAgentFleet
$ make doctor && make env && make up

    Then open http://localhost:8081, connect a model, launch an agent.

    SDK and fleetctl
    $ pip install open-agent-fleet
$ fleetctl host --root ~/projects

    Attach your own PC for sessions with Oaf and for the agents that live on it. It asks in your terminal before running a command or writing a file.

    Bring Claude Code
    $ fleetctl agent add Claude --kind claude_code \
    --folder ~/projects/app --reports-to Builder
$ fleetctl ticket new "Fix the login test" \
    --to Claude --review-by Checker

    Or --kind codex, hermes, openclaw, webhook. Also from the console: Org → Add agent.

    Support the project

    Keep the fleet flying.

    OpenAgentFleet is free and open source. If it is useful to you, you can support development with XRP or Bitcoin. Every bit goes into more agents, more testing and more hours on it.

    XRP

    Ripple
    tag required
    QR code for the XRP address rf82s1CDagppvM6ATqc1nSrL6GackzHJrm with destination tag 796343731
    Address
    rf82s1CDagppvM6ATqc1nSrL6GackzHJrm
    Destination tag / memo
    796343731

    The destination tag is required. Send XRP to this address with tag 796343731, or the transfer will not be credited. The QR code includes it.

    Bitcoin

    BTC
    no memo
    QR code for the Bitcoin address bc1qvre807vxh08puxwc2z5adnm59tta7v5mqmky45
    Address
    bc1qvre807vxh08puxwc2z5adnm59tta7v5mqmky45

    No memo or tag needed. Scan the code or paste the address into your wallet.

    Starring the repository and telling someone about it helps too.