One-click deploy

Pick a server option, wait for the first boot, then open the dashboard /setup link it gives you. /setup adds your harness credentials and integrations, so no option asks for them up front.

The default stack (a lead, 10 workers, agent-fs and TEI) needs at least 4 vCPU, 8 GB RAM and 40 GB of disk. Only the solo preset fits 2 vCPU / 4 GB. See Host requirements.

Cloud-init on any VPS

deploy/cloud-init.yaml turns a fresh Ubuntu 24.04 or Debian 12 server into a running swarm. On first boot it:

1. Sets Docker's default address for published ports to 127.0.0.1, then installs Docker Engine with Compose v2, Caddy and ufw.
2. Opens only ports 22, 80 and 443.
3. Points Caddy at <public-ip>.sslip.io (for example 203-0-113-7.sslip.io), or at your domain. Caddy gets the certificate from Let's Encrypt.
4. Runs install.sh --no-tunnel --public-url https://<host> --bind 127.0.0.1. The swarm sets PUBLIC_MCP_BASE_URL to the HTTPS address and publishes its ports on 127.0.0.1 only.
5. Writes the dashboard /setup link to /root/agent-swarm-setup-link.txt and shows it when root logs in.

To use it:

1. Create a server with at least 4 vCPU and 8 GB RAM on Ubuntu 24.04.
2. Paste the whole file into the user data field: Cloud config on Hetzner, User data on DigitalOcean, Vultr and Akamai, Launch script on Lightsail. Keep ports 80 and 443 open in the provider's firewall, if it has one.
3. Wait about 10 minutes, then ssh root@<server-ip>. The login message shows the /setup link. To follow the first boot: tail -f /var/log/agent-swarm-deploy.log.

Hetzner can fetch the file instead of a paste. Put this in Cloud config:

#include
https://raw.githubusercontent.com/desplega-ai/agent-swarm/main/deploy/cloud-init.yaml

Your own domain

Point an A record at the server, then set SWARM_DOMAIN near the top of the file before you paste it:

      SWARM_DOMAIN=swarm.example.com

The same block takes SWARM_PRESET (full by default) and any AGENT_SWARM_* installer variable, such as AGENT_SWARM_CLI_VERSION.

Security

• Docker-published ports skip ufw. Before the first container starts, the installer sets 127.0.0.1 as the default host address of the stack's Docker network (in docker-compose.override.yml), so every port the swarm publishes listens on loopback, whatever agent-swarm CLI release npm serves. Port 3013 is reachable only through Caddy. If a port still ends up on a public interface, the script stops the stack and fails instead of exposing it.
• The API key stays in root-only files: /root/agent-swarm/.env, the setup link file and /var/log/agent-swarm-deploy.log. The first boot never prints it to the console or to cloud-init's log, and the login message shows it to root only.
• Data lives in Docker volumes on the server's disk. They survive reboots and upgrades. Back them up with your provider's snapshots.

DigitalOcean 1-Click

Not listed yet. The image is built from deploy/digitalocean and runs the same first boot as cloud-init. Until the listing is live, create a Droplet (4 vCPU / 8 GB, $48/month) and paste the cloud-init file as its user data.

Railway template

The template runs the API, agent-fs, TEI, a lead and 2 coders, one volume each, and needs the Hobby plan. Railway serves the API over HTTPS. The full service list is in deploy/railway/README.md.

Try it for an hour on E2B

This is a trial, not a deployment. The swarm runs in E2B sandboxes, and everything in it, including the database, is deleted when the sandboxes stop: after one hour by default (--timeout-sec), and at most 1 hour on E2B's Hobby tier or 24 hours on Pro. You pay E2B for the sandbox time.

export E2B_API_KEY=e2b_...
bunx @desplega.ai/agent-swarm e2b start-stack --yes --swarm demo --workers 2 --api-key "$(openssl rand -hex 16)"
bunx @desplega.ai/agent-swarm e2b swarms info demo --reveal-key

swarms info --reveal-key prints the dashboard link with the API key filled in. See the CLI reference.