Skip to main content
Trinity
Updated Nov 17, 2026main@87e21ff
Getting Started/Deploying Trinity

Deploying Trinity

Trinity is the operating system for the AI-native company — open source, self-hosted, that you own, and model-agnostic: each agent runs on Claude Code, Gemini CLI or OpenAI Codex. It is open source under the Apache 2.0 license and runs on infrastructure you control: a cloud server, your own hardware, or your laptop. Your agents, credentials and data stay there.

Choose where to run it

Trinity 1.0.5 is on the DigitalOcean Marketplace and the AWS Marketplace

Both images have Docker, Caddy, a host firewall and the Trinity release baked in, so first boot pulls nothing. On DigitalOcean, pick a size from the sizing table before you create the Droplet. On AWS, subscribe on the listing and launch the AMI from the EC2 console — see the AMI guide.

40 Agents, One Instance — a DigitalOcean install from zero, live

Sep 2026

I Built a DevOps Agent That Deploys Other Agents

Apr 2026

Control My DGX Spark From Anywhere

Jan 2026

What you need

  • •A machine with 8 GB of RAM for a working fleet (4 GB runs the platform and one or two agents). Prebuilt images and the DigitalOcean paths are x86-64; on an ARM server, build from source. For a server, pick the size before you create it: start at 4 vCPU / 8 GB RAM / 80 GB disk and see Sizing Your Server for larger fleets.
  • •Docker with the Compose v2 plugin (docker compose, no hyphen), and Git. The DigitalOcean paths install both for you.
  • •A Claude credential for your agents: a subscription token from claude setup-token (Pro or Max) or an Anthropic API key. Connecting one is the one required step of Trinity's first-run setup, on first sign-in.
  • •Claude Code, to connect to your instance and deploy agents onto it.

DigitalOcean

Marketplace 1-Click (no terminal)

Trinity 1.0.5 is published on the DigitalOcean Marketplace. Create a Droplet from the Trinity image there, or search Trinity under Marketplace when you create a Droplet in the control panel (8 GB recommended, 4 GB minimum). First boot obtains a Let's Encrypt certificate for the Droplet's IP and starts Trinity from images baked into the snapshot, so it is serving over HTTPS about ninety seconds later. No admin account exists yet: open https://<droplet-ip> and the first visitor creates it with an email and password.

Claim it right after creating the Droplet: until then, anyone who finds the IP can. To skip that window, supply the password as cloud-init user-data when you create the Droplet.

From your terminal

With DigitalOcean's doctl signed in, one command asks for a region and a name, creates a 4 vCPU / 8 GB Droplet (about $48/month), installs Trinity, and prints an HTTPS address when it answers. As on the 1-Click, you open that address and create the admin account there, then connect Claude in the setup that opens.

$ bash <(curl -fsSL https://raw.githubusercontent.com/abilityai/trinity/v1.0.5/scripts/deploy/trinity-do-create.sh)
Step-by-step guide

Either way, the first-run setup opens with a Secure this instance step: point a domain at the Droplet and save it as the Public URL, and Trinity obtains a certificate for it on the first visit. A Cloudflare Tunnel can then take the server off the public internet (Public Access).

AWS

Marketplace AMI (nothing to install)

Trinity 1.0.5 is on the AWS Marketplace as an AMI. Subscribe, launch it on a large or xlarge instance (t3a.large and up) with a public IPv4 address and a security group open on 80 and 443, then open https://<public-ip>/ once first boot finishes a few minutes later. The setup form asks for the EC2 instance ID before it creates the admin, so someone who finds the IP cannot claim it. The software is free; you pay AWS for the instance.

Instance sizes, first boot and what to check →

Script install (your own instance)

On your own EC2 instance (x86_64, 8 GiB, Ubuntu 24.04, public IPv4, security group open on 80 and 443), as root:

git clone --depth 1 --branch <release-tag> https://github.com/abilityai/trinity.git /opt/trinity
cd /opt/trinity
./scripts/deploy/start.sh --provision --cloud aws --hosted --unattended

It installs Docker, Caddy with a certificate for the instance's IP, and the host firewall, then Trinity — the same result as the AMI, on an instance you already have, with the option to set the admin password up front (otherwise it takes the same instance-ID claim). On both paths a timer follows a changed public IP after a stop/start. Full procedure, the CloudFormation template, and recovery: Deploy on AWS.

Any Linux server

On a server, pull-only is the path you want: every platform image and the agent base image are published to GHCR on each release, so nothing is compiled on the box.

git clone https://github.com/abilityai/trinity.git && cd trinity
cp .env.example .env                     # set ADMIN_PASSWORD (12+ chars)
echo 'TRINITY_IMAGE_TAG=v0.9.5' >> .env  # pin a release
./scripts/deploy/start.sh --hosted --unattended

Pin a release rather than latest, which moves on every release. A fresh hosted install starts on the PostgreSQL service bundled in the hosted compose file; start.sh generates its password. The checkout must stay beside the compose file (it mounts ./config/*), and upgrades are a re-run of start.sh --hosted with a new TRINITY_IMAGE_TAG, not a bare docker compose pull. Put TLS in front before exposing it: a reverse proxy with a certificate, or a Cloudflare Tunnel. Plain HTTP on a public IP sends passwords in the clear.

Full details, TLS and tunnel choices, building from source, and every .env key: Single-Server Deployment. Want a walkthrough for creating the VM on Hetzner, AWS, GCP or Vultr? The Ops Agent's /provision skill guides you through it.

Or let Claude Code do the install: /trinity:deploy-new-instance from the trinity plugin installs Trinity on any server you can SSH into (or in local Docker) and sets up an Ops Agent to manage it.

Your own computer

With Docker Desktop running:

git clone https://github.com/abilityai/trinity.git
cd trinity
./scripts/deploy/start.sh --unattended

--unattended generates the admin password and prints it at the end; without it, start.sh asks you to set ADMIN_PASSWORD in .env first. It also generates every other secret (SECRET_KEY, CREDENTIAL_ENCRYPTION_KEY, AGENT_AUTH_SECRET, the Redis passwords) and writes them to .env. Never change CREDENTIAL_ENCRYPTION_KEY or AGENT_AUTH_SECRET afterwards: stored credentials become unreadable and running agents lose their tokens.

The first run builds the agent base image, which takes 5–10 minutes. When the backend reports healthy, open http://localhost and sign in as admin. If port 80 is taken, set FRONTEND_PORT=8090 (or any free port) in .env.

Details: Local Development. An AI coding agent can also drive the whole install from the runbook at docs/AGENT_INSTALL_GUIDE.md.

Connect Claude Code and deploy an agent

Once Trinity is running, from any agent directory in Claude Code:

# Once per machine: your instance URL, then an emailed sign-in code
/trinity:connect

# Per agent: deploy this directory to the instance
/trinity:onboard

/trinity:connect creates an MCP API key for you and writes .mcp.json pointing at <your instance URL>/mcp. The sign-in code only goes to addresses on the instance's whitelist (Settings → Access → Email Whitelist); if none arrives, add your email there. MCP keys are listed under Settings → MCP Keys. See Building Agents.

On a fresh install, turn on External MCP access under Settings → MCP Keys first: until an admin does, every client outside the Trinity server — Claude Code on the same machine included — is refused with external_mcp_access_off. See MCP Server.

All install paths

PathHowBest forGuide
DigitalOcean Marketplace 1-ClickCreate a Droplet from the Trinity imageThe quickest route to a server, and the only one with no terminal at any point: HTTPS at the Droplet's IP in about ninety seconds, and you claim the admin account in the browserSingle Server → 1-Click
DigitalOcean, from your terminaltrinity-do-create.sh (needs doctl)Same result as the 1-Click, created from your terminal instead of the control panelDeploy on DigitalOcean
AWS Marketplace AMISubscribe on the AWS Marketplace listing, launch the AMI from the EC2 consoleTrinity in your own AWS account and VPC with nothing to install: HTTPS at the instance's public IP, and you claim the admin account in the browser with the EC2 instance IDDeploy on AWS → Marketplace AMI
AWS EC2, from a shellstart.sh --provision --cloud aws --hosted --unattended on your own Ubuntu 24.04 instanceThe same provisioned install on an instance you already have, or where you want the admin password set up front; a CloudFormation template ships in the repoDeploy on AWS
Vultr Marketplace appDeploy the Trinity app from the Vultr Marketplace (opens in the Vultr console)Trinity on a Vultr server with nothing to install: the latest stable release is set up at first boot in about five minutes, with HTTPS at the server's IP, and you claim the admin account in the browser. 4 vCPUs and 8 GB RAM minimumSingle Server → Vultr
Vultr, from a shellstart.sh --provision --cloud vultr --hosted --unattended on your own Ubuntu 24.04 instanceThe same provisioned install on a Vultr instance you already have, or where you want a specific releaseSingle Server → Vultr
Server, prebuilt images./scripts/deploy/start.sh --hostedAny x86-64 Linux VM, serving in about two minutes with no on-box buildsSingle Server → Prebuilt images
Server, build from sourcedocker compose -f docker-compose.prod.yml up -dARM servers, custom patches, the enterprise overlaySingle Server → Build from source
Local, build from source./scripts/deploy/start.shYour own machine, trying Trinity, development with hot reloadLocal Development

All eight share one installer (scripts/deploy/start.sh), one .env contract, and one set of day-two procedures. Fresh server installs from prebuilt images run on a bundled PostgreSQL database; SQLite reached end-of-support on September 1, 2026 (see Single Server → Database backend).

Key URLs

ServiceLocalServer
Web UIhttp://localhosthttps://trinity.your-domain.com
MCP serverhttp://localhost:8080/mcphttps://trinity.your-domain.com/mcp — the web server proxies /mcp, so an install that exposes only 80/443 still serves MCP (port 8080 also works where it is reachable; a fresh install publishes it on 127.0.0.1only). Outside clients need External MCP access on (Settings → MCP Keys)
Backend API docshttp://localhost:8000/docshttp://your-server:8000/docs, where port 8000 is reachable (the DigitalOcean paths close it to the internet)

Running it day to day

# Stop (keeps agent containers; picks the right compose file)
./scripts/deploy/stop.sh

# Start (add --hosted on a prebuilt-image install)
./scripts/deploy/start.sh

# Health check after any change
./scripts/deploy/verify-platform.sh

Never stop Trinity with docker compose down. It removes the agent network, and every agent then loops trying to restart against a network that no longer exists. If it already happened, run docker compose up -d, docker rm -f the stale agent containers, and start them again from the UI. Their workspaces are kept.

  • •Upgrading: on a prebuilt-image install, set the new TRINITY_IMAGE_TAG, check out the matching tag, and re-run start.sh --hosted. On a source install, pull and rebuild the platform images. Upgrading has the full procedure with rollback.
  • •Backups: Trinity backs up its database nightly at 03:30 UTC, plus a copy at boot before any migration, on the same disk. Copy them off the machine for disaster recovery. See Backup and Restore.
  • •Monitoring: health probes, resource thresholds and recovery patterns are in Monitoring.
  • •Ops Agent: trinity-ops-public is a Claude Code agent that runs these operations for you, locally or over SSH. See Ops Agent.

Detailed deployment guides

DigitalOceanOne command from your terminal to an HTTPS Droplet: doctl and an optional domain; you connect Claude on first sign-in
AWSThe Marketplace AMI, the script install on your own EC2 instance, the instance-ID admin claim, the IP-refresh timer, the CloudFormation template
Single ServerLinux VPS: prebuilt images (--hosted) or build from source, the DigitalOcean 1-Click and installer script, the Vultr Marketplace app and script install, the database backend, every .env key and which compose file forwards it
Local DevelopmentDocker Desktop, dev compose, hot reload, what start.sh generates
Public AccessCloudflare Tunnel, TLS postures, webhook surface, Slack/Telegram/WhatsApp integrations, /mcp through the tunnel
Hardening a Marketplace InstallBare IP → domain → tunnel or private network, and how to verify each stage
UpgradingPre-flight → backup → rebuild or re-pull → restart → verify → rollback
Backup and RestoreAutomatic nightly backups, manual copies, restore procedure, PostgreSQL
MonitoringSix-probe health check, resource thresholds, recovery patterns
Ops AgentA Claude Code agent for day-to-day operations: health, logs, restarts, updates, rollbacks, and provisioning guides for other clouds