Deploy on your server
This page is for the admin who puts Agentome on the company's own server, so that people's agents run there and not on their computers. It covers what the server needs, a Docker Compose install with places under gVisor, the same on Kubernetes, the network the agents get, and what to check before a pilot. The details of each step are in Operations; every setting is in Configuration.
What the server needs
- Linux on x86-64 or arm64, inside the company's network, with Docker Engine 26 or newer and the Compose plugin, or a Kubernetes cluster.
- gVisor (
runsc), so that every place and every run has a kernel of its own instead of the host's. Install it from gvisor.dev. - Room for places. Each person who works with agents has a place, a container of their own, which takes at most 2 CPUs and 8 GB of memory while it is awake, and up to 20 GB of disk for its sign-in and working folders. A place sleeps when nothing has run for a while.
- A name and HTTPS for the server, such as
agentome.example.com: people open it in the browser, and agents reach the server's gateway and git proxy by it. - Outbound HTTPS from the server to the model vendors (
api.anthropic.com,platform.claude.com,claude.ai,api.openai.com,auth.openai.com,chatgpt.com) and toregistry.npmjs.org, from which a place installs the official Claude Code and Codex. Everything else the agents reach, your GitLab, Jira or package mirror, you open per project.
Install with Docker Compose
Once gVisor is registered with Docker (step 2 below), one command does steps 1 and 3: it asks for the domain and whether places run here, and turns gVisor on for them.
curl -fsSL https://agentome.ai/install.sh | sh -s -- --domain agentome.example.com --placesBy hand, or to see what it does:
Install the server as in Getting started, and give it HTTPS: with a public name, the
proxyprofile gets a certificate from Let's Encrypt (HTTPS with the proxy profile); inside the company's network, use the company's certificate (HTTPS on a local network, A company certificate authority).Register gVisor with Docker in
/etc/docker/daemon.jsonand restart Docker. The second runtime is for a run's own Docker (step 4), and only for it:{ "runtimes": { "runsc": { "path": "/usr/local/bin/runsc" }, "runsc-docker": { "path": "/usr/local/bin/runsc", "runtimeArgs": ["--net-raw", "--allow-packet-socket-write"] } } }Turn places on in
.env, beside the settings of step 1, and start them:COMPOSE_PROFILES=proxy,places AGENTOME_PLACE_LAUNCHER_URL=http://place-launcher:8080 AGENTOME_PLACE_RUNTIME=runscdocker compose up -dIf your projects' tests start containers (testcontainers, a compose stack), give each run a Docker of its own, which lives inside gVisor and goes with the run:
AGENTOME_PLACE_DOCKER_RUNTIME=runsc-dockerIf you want agents only on the server, and no person's computer to connect, add
AGENTOME_COMPUTERS=false.
With more than a dozen places, give Docker smaller address pools, which must not overlap the company's networks: see Places on the server.
Install on Kubernetes
The manifests are in deploy/kubernetes of the release, with the places in the places
component. In short:
- Install gVisor on the nodes and create its RuntimeClass,
gvisorsay. - Set the namespace,
AGENTOME_PUBLIC_URLand the image tag inkustomization.yaml, the certificate iningress.yaml, and create the Secret with the secret key and the database's address. - Take the
placescomponent and setAGENTOME_PLACE_RUNTIME=gvisor(the RuntimeClass's name). The cluster's network plugin must enforce NetworkPolicies: Calico and Cilium do. kubectl apply -k deploy/kubernetes, then open/setupwith the setup code.
Every step is in Kubernetes. On Kubernetes a run has no Docker of its own yet: a project's databases come from its devcontainer, as its services.
The network agents get
A place and its runs have no way out but the egress proxy, which lets each through to what it needs and nothing else:
| Who | Reaches |
|---|---|
| The CLI, in a run's CLI container | The model vendors and Agentome |
| Every command, in the run's sandbox | Agentome and the hosts the task's project allows |
| A run's own Docker | The hosts the project allows, for pulling images |
| The place itself, between runs | The model vendors and npm, to install and sign in the CLI |
| A project's services (a database say) | Nothing outside |
An admin opens hosts per project in its Network section: the company's GitLab, Jira, a package mirror. Add public package registries adds npm, PyPI, Maven, Go, NuGet and Docker Hub at once, for a project without a mirror. A host a run was refused shows there for a week, with Allow beside it. The proxy lets a host through on port 443 unless the list names another, connects only to the address it checked, and reaches the company's own networks only for a host listed by its full name.
The proxy also connects to the server's own public name, for the CLI's gateway. Make sure the
name resolves to an address the server's containers reach: on a network whose router does not
loop a public address back inside, give the name an internal record, or an alias on the
egress network (see Places on the server).
Keep it running
The install script's agentome command does these from any folder on the server:
- Updates:
agentome backup, thenagentome update; places are made anew from the new image when they next wake. See Updating. - Backups:
agentome backupwrites the database; keep the secret key apart from it. See Backing up and restoring. - A key that may have leaked:
agentome rotate-keyseals every secret under a new one. See Rotating the secret key. - Logs: Admin → Logs shows what the server, the egress proxy and the runs did.
Before a pilot
- The server answers on HTTPS by its name, and
/setupis done. - Admins sign in with a code from an authenticator app, which Agentome asks of every admin.
-
AGENTOME_PLACE_RUNTIMEis set, anddocker infolistsrunsc. - A person made their place, signed in with
claude(orcodex login --device-auth) in its terminal, and an agent finished a task in it. - Each project lists only the hosts it needs, and its code host connection reaches only its repositories.
- CI secrets are out of an agent branch's reach: protected variables on GitLab, environments on GitHub (see Integrations).
- Backups run on a schedule, and the secret key is kept apart from them.
How runs are kept apart from what they must not reach is on Security.