AI coworkers that run on your computer, not someone else's cloud.
Create named Agents, chat with them one-on-one or in a group, and give them a Linux desktop you can watch. They use the AI subscriptions you already have. Sensitive steps wait for your approval.
Public alpha v0.3.1. Not production-ready.
Apache-2.0.
Linux 0.3.1 · Windows 0.3.0 · macOS 0.3.1 (signed) · Product site · Security model
That screenshot is the Agent Computer: a separate Linux guest with Chromium,
Terminal and Files. It is never your host desktop. The app (Tauri) starts a
local Go controller (botd) that owns every conversation, memory and run
record. Nothing leaves your machine unless you wire it up to.
Nineteen seconds: the setup wizard where you pick a lead engine, a workspace chat, then a phone pairing with the desktop host. This is a storyboard with a voiceover, not a recording of a live session. It was cut before the rename, so the UI in it still reads OpenFleetBots.
Every Agent in the sidebar carries its engine badge and its current state. Andy is working. Cami needs approval.
A gated routine stops in the chat and waits for you. Allow or deny the run, or save an always-rule so the app stops asking. Denied runs stay in the thread with the decision recorded.
The Routines panel lists each schedule and when it next fires.
The commercial "AI teammate" products are cloud-hosted, bundle-priced, and opaque about what their bots do between your messages. OpenAgentFleet takes the opposite bet:
- Your machine. The controller, the chats and the memory all stay local.
- Your subscriptions. It drives Grok Build (default), Codex App Server, bundled OpenCode, or optional Pi. It never silently swaps providers.
- Your eyes on everything. Watch runs live, stop them, take over the desktop, approve sensitive actions. Agent-to-agent collaboration stays off until you switch it on per Agent. When it is on, handoffs happen in the chat where you can see them.
- Named Agents with their own chat and memory. Create as many as you want. Mention a teammate, or open a group chat.
- One Agent Computer per workspace. A Linux desktop (Ubuntu 24.04 by
default) that starts only when browser or desktop work needs it. Agents
take turns on that same PC; creating an Agent does not spawn a VM. Pi has
no Agent Computer. Each Agent references that computer (default ID
workspace, backenddocker) so a later per-agent computer would not require a new Agent identity. - Approvals with memory. Sensitive commands and file changes route
through
botd. Rules you save (this principal, this resource, this operation) stop the app from asking twice. - Teach → Skill → Routine. Show an Agent a task once and keep it.
- Tasks, routines and connected apps. One list of every Agent run with its
brief, its saved answer and its output files; a calendar for scheduled work;
and read-only GitHub through your local
ghlogin. See Workspace productivity.
Extra search connectors (Web Search Plus, Hound, Donsetch), Workers, Fleet Host, and remote Computer workers live under Advanced. The app detects Claude Code, Codex CLI and Cursor as future or bounded adapters. None of them stands in for the workspace engine.
| Host | What to install |
|---|---|
| Linux | Unsigned .deb / .rpm / .AppImage from v0.3.1-alpha. Docker Engine recommended for the Agent Computer; the installer does not start it. |
| Windows | Unsigned NSIS is still v0.3.0-alpha. 0.3.1 Windows is not on the tag yet. Docker Desktop recommended. Computer View on Docker Desktop WSL2 failed live QA 2026-08-25 (daemon/build). Docker Desktop.exe must be open as the signed-in user; a service in Running is not enough. |
| macOS (Apple Silicon) | Signed v0.3.1-alpha DMG, SHA-256 570731b921b64da9ee2d94d35467b6388372b48146e8270eb1642c9309209f02. Intel Macs are not supported. |
Checksums live on the release page. Release runbooks: Linux · Windows · macOS.
- Pick an engine. Grok Build is the default.
- Create an Agent and chat.
- When a task needs a browser or desktop, start the Agent Computer. Opening the app never starts Docker or a VM on its own.
- Create a second Agent and put both in a group chat.
iOS and Android pair with the desktop host over private Tailscale. The
phone is a remote control, not a second runtime. From the phone you can chat,
approve, stop a run, and watch the computer. Typing passwords and driving the
full desktop stay on the host. No store builds yet; see mobile/.
Provider CLIs run on the host. The Agent Computer isolates browser and desktop work, not those processes.
- Grok Build and Codex App Server route sensitive commands and file
changes through
botdapprovals. - OpenCode keeps its own
askmode; its dangerous auto mode stays off. - Pi signs in with
pi /login(~/.pi). Picking a Grok or GPT model id inside Pi still runs through Pi.
One Linux desktop with Chromium, Xfce, Terminal and Files. Ubuntu 24.04 is the default; Ubuntu 26.04 and Debian 13 are optional. Default budget: 4 CPU, 4 GiB RAM, 25 GiB disk. It starts only when you open Computer View or an approved task needs it.
Backends: Docker Engine (Linux), Docker Desktop (Windows; Computer View on WSL2 failed live QA 2026-08-25, not a working session), Colima recommended on macOS (Docker Desktop / OrbStack also work).
No host root mount. No Docker socket in the guest. Details: Agent Computer backends · FAQ.
Requires Go, Node.js + pnpm, Rust, uv/uvx, and OpenCode 1.18.10.
git clone https://github.com/robbyczgw-cla/openagentfleet.git
cd openagentfleet
go test ./...
cd client
pnpm install
pnpm run prepare:sidecar
pnpm run tauri devBrowser-only UI: cd client && pnpm install && pnpm dev with
VITE_BOTD_URL set. Useful for development; not a replacement for the
packaged app.
Apache-2.0 · Runtime notices: NOTICE.md · Docs map: docs/README.md