Self-host guide

Set Sayabot up with your agent.

Cursor, Claude, Codex, and Gemini install it for you. Paste the prompt into one of them. They clone the repo, run the installer, and check that your server answers. Sayabot's own agents live on that server after it is up.

Setup prompt copied

Copies the setup prompt. Your agent installs Sayabot. It does not sign in as a Sayabot agent.

On a VPS clone

What actually comes up

Checked against the installer and the compose files. Ready means the first install starts it. Extra step means the server runs without it until you add that setting.

  • Ready

    Web workspace, rooms, schedules, memory

    Started by ./deploy/install.sh together with nginx and the orchestrator.

  • Ready

    Sign-in on this server

    Bundled Supabase serves Postgres, auth, realtime, and file storage on your domain. Email delivery and Google or GitHub sign-in need SMTP or OAuth settings.

  • Ready

    Turns, canvas, MCP, approval gates

    Turns use the room lock. Canvas is wss://DOMAIN/collab. MCP is https://DOMAIN/api/v1/mcp. Dangerous tools wait for a yes.

  • Ready

    Sealed secrets

    install.sh generates SAYABOT_KEK. Saved keys are encrypted with it. The file must stay mode 600, and the key must never change.

  • Ready

    Mac and phone apps

    Point them at https://DOMAIN. They read /api/instance first and show your server's name.

  • Extra step

    Agent computers and egress

    Default services are orchestrator and web. Run setup-computers.sh, set SAYABOT_APP_SERVICES to include computerd and egress, then deploy again.

  • Extra step

    A model for the agents

    Add a provider under AI accounts, or set SAYA_AI_* and run ./deploy/deploy.sh reload.

  • Off

    Plans, limits, and billing

    Those exist only when SAYABOT_EDITION=cloud. A server you host does not run them.

Four steps

Install

The same commands the copied prompt tells your agent to run. You can run them yourself.

  1. 01

    Prepare the server

    Ubuntu 22.04 or 24.04, or Debian 12. Two vCPU, 4 GB RAM, 40 GB disk. Point a domain at the machine, or have a Cloudflare Tunnel token ready.

    terminal
    scp deploy/scripts/setup-vps.sh root@YOUR_IP:/root/
    ssh root@YOUR_IP 'bash /root/setup-vps.sh --user deploy --ssh-key "$(cat ~/.ssh/id_ed25519.pub)"'

    Confirm ssh deploy@YOUR_IP works before you close the root session. The script turns off password login.

  2. 02

    Clone and install

    The installer writes deploy/.env.production, generates every secret, gets a certificate or uses your tunnel, and starts Sayabot. Re-running it keeps SAYABOT_KEK.

    terminal
    git clone https://github.com/anurag3407/sayabot.git ~/sayabot && cd ~/sayabot
    ./deploy/install.sh

    Choose “Everything on this server” unless you already have a Supabase project. The database URL must use port 5432.

  3. 03

    Turn on agent computers

    The first install starts the web app and the orchestrator. Computers and the outbound proxy are a second step, because the computer image is heavy.

    terminal
    sudo ./deploy/scripts/setup-computers.sh
    # paste the lines it prints into deploy/.env.production, including:
    # SAYABOT_APP_SERVICES="orchestrator web computerd egress"
    ./deploy/deploy.sh

    A host with /dev/kvm gets a microVM per agent. Any other VPS gets gVisor. A tiny machine can use a cloud computer with COMPUTER_BACKEND=e2b instead.

  4. 04

    Open it and connect a model

    Create the first account at your domain. Sayabot's own agents need a model before a turn can run. Add one under AI accounts, or set SAYA_AI_* for everyone on the server.

    terminal
    ./deploy/deploy.sh reload    # after SAYA_AI_* changes
    ./deploy/deploy.sh status
    ./deploy/deploy.sh backup     # bundled database

    Mac and phone: open https://YOUR_DOMAIN/download, or choose your server in the app. The apps check /api/instance and show your server's name.

After it boots

Checks your agent should run

Each line is a pass or a fail. The prompt asks for a report, and it tells the agent not to print keys.

  • Healthcurl -fsS https://DOMAIN/api/health returns ok: true and service: sayabot-web.
  • This servercurl -fsS https://DOMAIN/api/instance returns sayabot: true and edition: self-hosted.
  • Containers./deploy/deploy.sh status shows web, orchestrator, and nginx healthy. After step 03, computerd and egress too.
  • Secretsdeploy/.env.production is mode 600. DATABASE_URL uses port 5432. SAYABOT_KEK is set, and never printed or changed.
  • A turnThe login page renders. A room starts a turn once a model is connected.
  • A yesA dangerous tool pauses for approval. The canvas socket upgrades at /collab.
  • MCPhttps://DOMAIN/api/v1/mcp answers. A bundled database backup lands in deploy/backups.
  • Day to daygit pull && ./deploy/deploy.sh, then status, logs, backup, and rollback.
Hand this over

The prompt

Same text the button copies. Paste it into Cursor, Claude, Codex, or Gemini on the machine that will run Sayabot.

setup-prompt.txt
Set up Sayabot on this Linux server from the public repository, then verify the self-hosted stack. Follow the installer in the clone. Do not invent a deploy path, and do not print secrets.

Repository: https://github.com/anurag3407/sayabot
License: AGPL-3.0. Keep SAYABOT_EDITION=self-hosted. A self-hosted server has no plans, no usage limits, and no billing. Do not set the edition to cloud, and do not run anything under ee/.

What the installer starts
- nginx, the Next.js web app, and the orchestrator
- Let's Encrypt on ports 80/443, or a Cloudflare Tunnel when a tunnel token is provided
- Postgres, sign-in, realtime, and file storage on this machine when database mode is bundled (the recommended answer)

What the installer does not start
- Agent computers (computerd) and the SSRF-safe egress proxy. They start only after deploy/scripts/setup-computers.sh and SAYABOT_APP_SERVICES includes computerd and egress. The default value is "orchestrator web".

Requirements
- Ubuntu 22.04 or 24.04, or Debian 12
- 2 vCPU, 4 GB RAM, 40 GB disk. The first web build needs about 3 GB; setup-vps.sh adds 4 GB of swap.
- A domain whose A record points at this server, or a Cloudflare Tunnel token

Install
1. From a laptop, once, as root:
   scp deploy/scripts/setup-vps.sh root@SERVER:/root/
   ssh root@SERVER 'bash /root/setup-vps.sh --user deploy --ssh-key "$(cat ~/.ssh/id_ed25519.pub)"'
   Confirm ssh deploy@SERVER works before closing the root session. This turns off password login.
2. As the deploy user:
   git clone https://github.com/anurag3407/sayabot.git ~/sayabot && cd ~/sayabot
   ./deploy/install.sh
   Answers: the domain, the name shown in the apps, HTTPS (Let's Encrypt, or a Cloudflare Tunnel token), and the database. Choose "Everything on this server" unless an existing Supabase project was provided. The database URL must use the session pooler, port 5432, never 6543.
3. Turn on agent computers. As root:
   sudo ./deploy/scripts/setup-computers.sh
   It picks a microVM when /dev/kvm exists, otherwise gVisor. Paste the lines it prints into deploy/.env.production. They set COMPUTER_BACKEND and:
   SAYABOT_APP_SERVICES="orchestrator web computerd egress"
   Then, as the deploy user: ./deploy/deploy.sh
4. Open https://DOMAIN and create the first account. Connect a model under AI accounts. To share one model with every user, set SAYA_AI_* in deploy/.env.production and run ./deploy/deploy.sh reload.

Verify each item and report pass or fail. Do not print SAYABOT_KEK, database passwords, or API keys.
- curl -fsS https://DOMAIN/api/health returns JSON with "ok": true and "service": "sayabot-web"
- curl -fsS https://DOMAIN/api/instance returns "sayabot": true and "edition": "self-hosted", plus the server name
- ./deploy/deploy.sh status shows web, orchestrator, and nginx healthy. After the computer step, computerd and egress are healthy too.
- deploy/.env.production is mode 600 and SAYABOT_KEK is non-empty
- DATABASE_URL contains :5432 and does not contain :6543
- https://DOMAIN/login renders
- After a model is connected, a room accepts a message and an agent turn starts
- A tool marked dangerous pauses for a human yes and does not run on its own
- The canvas socket at wss://DOMAIN/collab accepts a WebSocket upgrade
- MCP is reachable at https://DOMAIN/api/v1/mcp
- The Mac and phone apps can read this server from https://DOMAIN/api/instance before anyone signs in
- When the database is bundled, ./deploy/deploy.sh backup writes a dump under deploy/backups

Never change SAYABOT_KEK after it is written. Losing it makes every saved key unreadable. Copy deploy/.env.production off the server.

Day to day
- git pull && ./deploy/deploy.sh
- ./deploy/deploy.sh status
- ./deploy/deploy.sh logs web
- ./deploy/deploy.sh rollback

If a check fails, read docs/self-hosting-quickstart.md and deploy/README.md in the clone. Do not open the orchestrator's /internal API, do not weaken TLS, and do not commit deploy/.env.production.

Longer manuals live in the repository: the full self-hosting guide and the ten-minute quickstart.