Expert

The supported interfaces are WhatsApp and hello.expert. The native iOS app was removed; the web workspace works on mobile.

Sign in at hello.expert to watch the desk and talk to the agent. One Fly Machine per tenant: Blode stays mblode-computer; Vibey is a second computer (vcmc-computer).

git clone https://github.com/mblode/expert && cd expert && npm install

Requires Node 24 and a running Docker daemon (Docker Desktop, OrbStack, colima). Without Docker the hub runs against a fake desk so you can still pair and poke around.

# builds the desk container, starts the agent and the hub, prints a pairing QR
npm run up

# in a second terminal: the web client on http://localhost:3000
npm run web

Open the web client, sign in with an email code, and the desk appears: a 1280x800 Debian desktop with Chromium, streamed view-only. Give the agent something to do in the chat, and take the seat when it asks.

Provision more Bots, each on its own screen of the same box:

npm run bot -- new night   # Bot night is live on screen 2. token: bot_…
npm run bot -- ls          # main  screen 1  AGENT / night  screen 2  AGENT
npm run bot -- rm night    # frees the screen

• Watch the agent work: the screen streams live over VNC; the model sees screenshots, clicks by pixel, types, scrolls, zooms.
• Take the seat: one click makes the screen yours, with a local cursor, keyboard, and clipboard both ways. The agent's next action waits until you tap I'm done.
• Let it ask: passwords, 2FA, captchas and payments are the human's job. The agent hands the desk over, or asks for a masked secret that goes straight to the box clipboard and never reaches the model.
• Run several Bots: one shared box, one screen per Bot, up to sixteen. Bots share files and browser sessions and are not security boundaries.
• Keep your files: /workspace and the Chromium profiles survive restarts and image rebuilds, the same boundary Grok Bot draws. Installed packages do not; keep the list in /workspace/packages.md.

Step by step, with the check that proves each step worked: docs/DEPLOY.md. The shape of it:

The cloud path is one Fly Machine per tenant in syd running the desk, one agent process per Bot, and the hub. There is one tenant today, Vibey, in fly.toml; a second is another app on the same image (docs/plans/gateway.md).

fly secrets set COMPUTER_SETUP_CODE="$(openssl rand -hex 16)"
fly secrets set AI_GATEWAY_API_KEY="…"
fly deploy

It suspends when idle: CPU and RAM go to $0; the volume still bills. Do not create a second volume. What makes the computer Vibey is the tenant content on its volume under /workspace/.bots/main/data, not a different build.

The product web is apps/web on Vercel with Root Directory apps/web. It is the control plane: a signed-in user is bound to a computer. Required variables:

A WhatsApp tap opens a short-lived invite: /desk/<token> is the phone desk (take/yield seat, pointer, keyboard). /plugins/<token> adds an Eve connection file under agent/connections/ on the guest. Plugins are files, not a database table. Skills stay as files too.

Push the schema once with cd apps/web && npx drizzle-kit push. An always-on VPS is the alternative: deploy/cloud-init.yaml.

• The protocol is small on purpose: five model tools (send_message, computer, shell, read_file, write_file) and a Seat service for humans. Clipboard, the VNC URL and the pointer are never model tools. api/DESIGN.md is the contract; api/RESEARCH.md is the argument.
• This is a clean-room clone of Grok Bot's computer: what it is, how far this is from it, and the roadmap are in docs/GROK-BOT.md. The current engineering audit is docs/AUDIT.md.
• Agents working on this repo read AGENTS.md: commands, layering rules, and the gotchas.

Apache-2.0