First public alpha. Self-hosted AI agent workspace: each agent gets an isolated Linux desktop, you watch it live, and you can take over at any time.
This is a source release. There is no pre-built binary; `make up` builds the API, supervisor, and desktop images from this tree.
### Highlights
- **A real computer per agent** — Debian + XFCE + Chromium in Docker. Browser, terminal, files, and GUI, with a live noVNC view.
- **Cua desktop driver** — clicks, typing, windows, clipboard, and the visible terminal go through Cua (`cua-driver-rs` v0.23.2). The old AT-SPI / CDP path is gone.
- **Chat that carries work** — text, images, and file attachments; the agent opens them on its own desktop. Inbox copies expire on their own.
- **Take over, then hand back** — sign in, pass a check, or nudge the same desktop by hand. Saved logins live in an encrypted vault and never pass through the model.
- **Several agents and groups** — shared Team computers or private desktops. `@name` wakes one agent; otherwise a short router picks who should answer.
- **Teach by demo, then schedule** — turn a walkthrough into a skill; cron runs it again.
- **Your models and tools** — xAI, OpenCode Go, OpenAI-compatible endpoints, MCP, and file skills.
- **Voice calls** — talk to the agent after you enable a voice provider.
- **Desktop and phone UI** — English and Traditional Chinese. On a phone: collapsible chat, keyboard, trackpad, right-click, drag, and scroll.
- **Runs that finish the job** — no hard turn quota on real work. The agent has to verify against the current screen before it stops; looping is paused for you instead of being cut off. Hover the thinking avatar for the live trace; a failed run can be retried.
- **Memory and hygiene** — optional long-term memory (pgvector), room-scoped recall, and retention jobs for traces, recordings, and checkpoints.
### Install
You need Docker and Compose, Git, Make, Python 3, and an API key for a supported model.
Edit `.env` and set one of `XAI_API_KEY`, `OPENCODE_GO_API_KEY`, or `OPENAI_API_KEY`. Then:
```bash
make up
make health
```
Open [http://127.0.0.1:3101](http://127.0.0.1:3101) and sign in with `LAZYBOY_APP_TOKEN`. The first run builds images from source and takes a while.
### Alpha caveats
- Early software. Expect rough edges in the desktop driver, group routing, and long-running tasks.
- The API binds `127.0.0.1` by default. For a phone on your network set `LAZYBOY_BIND_IP=0.0.0.0` (token must be at least 32 characters). Put it behind HTTPS before you expose it.
- Do not publish supervisor port `7091` or mount the host Docker socket into an agent computer.
- You bring the model API key. Screenshots and page contents can still reach the model; keep passwords in the vault, not in the prompt.
See [Operations](./docs/operations.md) for resource limits, environment variables, and the security model.
### 繁體中文
第一個公開 alpha。自架 AI Agent 工作空間:每個 Agent 有自己的隔離 Linux 桌面,你可以即時觀看,也可以隨時接手。