Requirements

Runtime

Item	Version	What for
Node.js	≥ 20	Building and testing the packages. The repository’s engines field requires it.
Docker	any recent	Running the runner in a container. The image is node:22-alpine.
npm	10+	The repository’s workspaces.

You don’t need Node installed to run the containerized runner: only to build the packages or develop connectors.

A control-plane on the other side

The Agent is the “hands” half. On its own it boots, connects, and waits: with no control-plane listening on the tunnel, it does nothing useful. Before installing, have these ready:

• The image address and pull authorization for your account or project.
• The tunnel URL (RUNNER_TUNNEL_URL), in production: wss://tunnel.rootpilot.sh:8443.
• Your tenant certificate: runner.crt, runner.key, and ca.pem.

All three are delivered by RootPilot when your environment is opened.

The hostname matters even when another one resolves to the same IP

The tunnel certificate carries SAN DNS:tunnel.rootpilot.sh. Connecting through any other hostname fails TLS hostname verification with is not in the cert's altnames. An open port is not the same as the right hostname. The certificate’s SAN decides.

Egress

The runner needs egress, and only egress. No inbound ports.

Always:

• tunnel.rootpilot.sh:8443 (WSS + mTLS), the tunnel.
• The registry the image comes from (ECR or Artifact Registry), on pull.

No enrollment, no app.rootpilot.sh

On an image install the Agent receives a ready-made certificate and does not enroll, so it needs no egress to app.rootpilot.sh:443. That rule is only required if you use the automatic enrollment path — see Enrollment.

Depending on which connectors you enable (only the ones you use):

Connector	Egress host
datadog	api.datadoghq.com
github	api.github.com
aws, kubernetes	*.amazonaws.com, *.eks.amazonaws.com
gcp	*.googleapis.com
amplitude	amplitude.com
slack	slack.com
cloudflare	api.cloudflare.com
azion	api.azion.com
vercel	api.vercel.com
checkly	api.checklyhq.com
grafana	your datasource endpoints (Mimir, Loki, Tempo, stack)
mongo, sonarqube, jira	your own hosts, which you supply in configuration

That list is declaratory, not a firewall

Native connectors wrap third-party SDKs that make their own HTTP calls, so the egress allowlist declared in a manifest is not enforced for them: it documents intent. Real egress mediation is a property of the sidecar mechanism. If you need a guarantee, enforce it in your own egress controls: security group, NetworkPolicy, or proxy.

Sizing

This repository publishes no CPU or memory figures, because there is no versioned benchmark here, and inventing one would be worse than omitting it. What the construction does let us state:

• The process is a single-container Node process, with no worker pool.
• The SQLite cache is :memory: by default, so memory grows with the volume of state learned during the session and resets on every boot.
• RUNNER_IDEMPOTENCY_MAX_ENTRIES (default 2000) bounds the idempotency cache on purpose, so a long-lived runner doesn’t grow without a ceiling.

Start small, observe, adjust. The runner is cattle: resizing means bringing up another one.

Credentials for your sources

None are required to boot. Every connector is lazy, instantiated on first use, so a missing key only makes those operations answer auth_error; everything else keeps working. The full list is in Connector credentials.

The Profiler has its own prerequisites, and harder ones

Everything above is about the Agent. The RootPilot Profiler is a second artifact — optional and privileged — and what it demands of the host does not overlap with this page: kernel ≥ 5.8 with BTF compiled in, cgroup v2, tracefs mounted, and five kernel capabilities.

Those gates reject hosts nobody expects to see rejected, so they are checked before the POC rather than during it: curl -O https://docs.rootpilot.sh/profiler-preflight.sh && sh profiler-preflight.sh. The script is published in full on the Profiler page, so it can be read before it is run.