Deployment

To keep teammates working while your laptop is closed, run open-bot in the cloud. This page describes the shape of a deployment and walks through it on Railway, which the project was built against. The same shape works on any host that runs containers with a private network and a persistent volume.

Services

The three services built from the repository use its root as their root directory, because they need the shared workspace packages and the root lockfile.

1. Secrets

Generate each secret separately:

openssl rand -hex 32      # BETTER_AUTH_SECRET, WORKER_SECRET, BOTD_TOKEN (one each)
openssl rand -base64 32   # ENCRYPTION_KEY: exactly 32 bytes, base64

Back up ENCRYPTION_KEY. Every stored API key, vault item and connection secret depends on it.

2. Postgres

The 0020 migration runs CREATE EXTENSION IF NOT EXISTS vector, so the database must have the pgvector extension available. Use a Postgres image or template that includes it, such as pgvector/pgvector:pg17, which is what local development uses.

3. The computer

• Build from images/computer/Dockerfile (on Railway, set RAILWAY_DOCKERFILE_PATH=images/computer/Dockerfile), and use the image's own entrypoint.
• Mount a persistent volume at /home. A Dockerfile VOLUME is not enough.
• Give it no public address. botd listens on port 7000.
• Health check /healthz, with up to 300 seconds to start.

BOTD_TOKEN=<same as the worker>
MAX_SCREENS=2                  # start low, measure memory, then raise
RAILWAY_RUN_UID=0              # botd must start as root to manage its users
RAILWAY_DEPLOYMENT_DRAINING_SECONDS=30
RAILWAY_SHM_SIZE_BYTES=2147483648   # 2 GB of shared memory for the browsers

Budget memory for the desktop, Codex, and each running screen. Measured locally, an idle computer uses about 370 MB, and each blank screen about 330 MB more.

4. The worker

pnpm --filter @open-bot/worker build    # esbuild bundles the worker and its workspace packages
pnpm --filter @open-bot/worker start    # node dist/worker.mjs

Run migrations before the worker starts, as a pre-deploy step on the worker only:

pnpm db:migrate

The migration script runs through tsx, so keep development dependencies installed (RAILPACK_PRUNE_DEPS=false on Railway).

NODE_ENV=production
WORKER_PORT=4001
DATABASE_URL=<private Postgres URL>
BOTD_URL=http://<computer private host>:7000
BOTD_TOKEN=<same as the computer>
WORKER_SECRET=<same as web>
ENCRYPTION_KEY=<same as web>
APP_URL=https://<web domain>            # webhook URLs start with it
WATCH_ALLOW_PRIVATE_URLS=false
MAX_CONCURRENT_RUNS=2
ELECTRON_SKIP_BINARY_DOWNLOAD=1          # the workspace includes the desktop app

Give the worker a public domain on port 4001. Browsers connect to its /vnc endpoint and the desktop app to /egress/agent directly, with short-lived signed tokens. Every other route needs WORKER_SECRET.

Turn off any "sleep when idle" feature for the worker and the computer, so schedules, polling and sessions keep running.

5. The web app

pnpm --filter @open-bot/web build
pnpm --filter @open-bot/web start       # next start on port 3000

NODE_ENV=production
OWNER_EMAIL=<your email>
DATABASE_URL=<private Postgres URL>
BETTER_AUTH_SECRET=<random>
BETTER_AUTH_URL=https://<web domain>
WORKER_URL=http://<worker private host>:4001
WORKER_SECRET=<same as worker>
ENCRYPTION_KEY=<same as worker>
NEXT_PUBLIC_WORKER_WS_URL=wss://<worker domain>
ELECTRON_SKIP_BINARY_DOWNLOAD=1

NEXT_PUBLIC_WORKER_WS_URL is built into the client, so rebuild after changing it. Use the targeted build commands above rather than the root pnpm build, which also builds the desktop app.

6. First sign-in and smoke test

1. Deploy Postgres and the computer, then the worker (with migrations), then the web app.
2. Open the web app and create the owner account with OWNER_EMAIL.
3. In Settings → ChatGPT, choose Sign in on the computer, then complete the device-code login. The login is saved on the computer's volume.
4. Optionally add an OpenAI API key and a monthly cap.
5. Create a teammate. Open its screen, load a page, take over, hand back, and send a message.
6. Restart the computer and check that files, the ChatGPT login and browser sign-ins survived.
7. Run the desktop app against the deployment with OPEN_BOT_URL=https://<web domain>, sign in inside it, and check that the computer panel shows browsing through your Mac.

Backups

Not built yet

• There is no provider integration that creates the computer for you. You deploy the image as a service.
• The desktop app has no signed macOS build. pnpm --filter @open-bot/desktop package builds an unsigned one.