OpenClaw Deployment (Provisioning Flow)
Last updated: 2026-08-06
This document describes how an OpenClaw runtime is provisioned on the HostAgentics platform: the service layout, persistence, gateway credential, domain, health verification, hardening, and channel setup. It is an internal engineering document; customer-facing claims must stay within what this flow actually delivers.
1. Service layout
An OpenClaw runtime is a **dedicated gateway service** from the official OpenClaw production image (`ghcr.io/openclaw/openclaw:2026.7.1-2` at the tested default, digest-pinned via the version catalog). One gateway cell per tenant matches the official multi-tenant model. The service runs in the runtime's shard project and region.
2. Persistence: config and workspace volumes
Both directories live on the runtime's own persistent volume; storage is never shared between runtimes, and the volume is attached to exactly this runtime's service.
3. Gateway credential: generated, encrypted, per runtime
Each OpenClaw runtime gets a **generated gateway token** (`gateway-token` → `OPENCLAW_GATEWAY_TOKEN`), created at provisioning time, stored in the runtime's encrypted secret scope (AES-256-GCM envelope, `docs/secret-rotation.md`), and injected into the service environment at configure time. There is no shared platform credential: one token per runtime, and a rotated token stops working on redeploy. Model access is provided via a restricted per-runtime key (HostAgentics credit) or the customer's BYOK key — never the platform management key.
4. Domain and HTTPS
The runtime is assigned a branded domain `<subdomain>.runtime.hostagentics.com` with automatic HTTPS: the provider assigns the domain, DNS records are confirmed, and certificate status is polled until `active` before the runtime is presented as ready. The gateway's base URL is derived from that domain.
5. Health verification
The gateway listens on port **18789**. The health probe sends an authenticated request to the gateway root; **any HTTP response — including a 401 auth challenge — proves the process is live and enforcing authentication**, and counts as healthy. A gateway that answers without an auth challenge is treated as unhealthy (it is not enforcing token auth). Provisioning cannot reach `running`, and restores cannot be marked completed, until this probe passes.
6. Hardening (no host access)
OpenClaw runtimes run with the platform's standard sandbox posture (see `docs/sandboxing.md`):
7. Channel setup
OpenClaw connects to the customer's messaging channels through guided setup: **web chat, Telegram, Discord, Slack, and WhatsApp**. Channel credentials (bot tokens, webhook secrets) are submitted once, stored encrypted in the runtime's secret scope, and never displayed again in plaintext. The platform does not read agent conversation content.