Google Cloud Run instances target always-on personal agents

Google Cloud has introduced Cloud Run instances, a preview primitive for dedicated, singleton compute aimed at long-lived personal AI agents such as OpenClaw and Hermes.

In a blog post dated August 27, 2026, product manager Ryan Pei and software engineer Matthew Robertson argue that those agents have infrastructure needs that sit awkwardly between two existing options. Cloud Run services scale to zero when requests stop, which is a poor fit for an agent that expects exactly one copy to keep running. A dedicated virtual machine, they write, means paying for full compute around the clock, managing operating system updates, opening firewall ports, and provisioning HTTPS endpoints yourself.

Cloud Run instances are Google’s answer: one replica, no autoscaling, a stable HTTPS URL, and the ability to stop and resume.

The launch post states that a Cloud Run instance with 1 vCPU and 1 GiB of memory running continuously for 30 days costs $5.70. Instances use shared vCPU with vCPU burst budgets, which Google positions as a predictable price for agents that are not doing compute-intensive work all the time and only spike when asked to perform a task. Continuous runtime is listed as up to seven days, with an automatic restart policy configured by default. The HTTPS URL is described as remaining unchanged across updates and restarts.

That combination is the product bet: always-on enough for a personal agent, cheap enough to leave running, and managed enough that you are not babysitting a VM.

OpenClaw, from laptop to a public URL

Google’s worked example is OpenClaw, an open-source personal AI agent that can perform tasks on a user’s behalf and, in Google’s words, become a better assistant over time. Many OpenClaw users start on a laptop until they need somewhere that will not shut down when the lid closes.

After OpenClaw configuration files are uploaded to a Cloud Storage bucket, the launch post deploys the agent with one command: gcloud beta run instances create, pointing at ghcr.io/openclaw/openclaw:latest, port 18789, a public URL, a Cloud Storage volume mounted at /home/node/.openclaw, and environment variables for OPENCLAW_GATEWAY_PASSWORD and GEMINI_API_KEY. Google says you can then keep the instance running as long as you want and interact with it over Telegram, WhatsApp, or another channel.

Hermes is named in the same breath as OpenClaw as an agent intended for individual developers or personal use. The post does not ship a Hermes deploy command.

A Codelab last updated August 27, 2026 fills in the production-shaped details the one-liner skips. It has you enable the Cloud Run, Secret Manager, Cloud Storage, Gemini, and Compute APIs; create a dedicated openclaw-sa service account; store the Gemini API key and a generated gateway password in Secret Manager; write an openclaw.json that turns on the Control Web UI on port 18789; and put that config in a Cloud Storage bucket. Deploy uses gcloud beta run instances deploy with 4 vCPU, 4 GiB of memory, --public, --sandbox-launcher, the same Cloud Storage mount (uid 1000, file and directory mode 0700), and --set-secrets rather than plaintext env vars.

The sandbox flag is there because OpenClaw’s default sandbox wants a local Docker daemon, which Cloud Run container images do not provide. With --sandbox-launcher, the Codelab says the platform injects a gVisor micro-VM sandbox binary at /usr/local/gcp/bin/sandbox for isolated code execution in about 25 milliseconds. Optional steps cover Telegram or WhatsApp, a Cloud Run sandbox-provider plugin uploaded next to openclaw.json, custom skills in the bucket, and signing into the built-in Web UI with the gateway password.

The CLI lifecycle, without the agent

If you just want the primitive, the create-instance quickstart (last updated August 26, 2026) walks the same beta CLI against Google’s hello container. Install the gcloud beta component, enable the Cloud Run Admin API and Cloud Logging, then gcloud beta run instances create with --image us-docker.pkg.dev/cloudrun/container/hello, a region, port 8080, and --no-invoker-iam-check so the URL is reachable without IAM authentication. On success the CLI prints a HTTPS URL on *.run.app. Logs live under resource.type="cloud_run_instance". Delete with gcloud beta run instances delete. The guide asks for Cloud Run Admin, Service Account User, and Logs Viewer.

Preview caveats

Cloud Run instances are in preview. Google says SSH access for both instances and services is coming soon. The $5.70 figure is Google’s stated price for the 1 vCPU / 1 GiB size running 30 days; the OpenClaw Codelab’s 4 vCPU / 4 GiB shape is a different size, and Google does not quote a monthly number for it in these pages.

OffDeal, which Google describes as an AI-powered investment bank for small businesses, is already running a long-lived agent on instances. Luis Ruiz Morel, Member of Technical Staff at OffDeal, is quoted: “We're currently using Cloud Run instances as our primary infrastructure for our long-running agent. It reduced cold starts by 88%. Everything was very straightforward to implement, and it has been very reliable.”

For practitioners who have been parking OpenClaw or Hermes on a laptop, the preview is a managed always-on HTTPS URL without a full VM, at a stated $5.70 a month if you stay on 1 vCPU and 1 GiB.

Work through Google’s OpenClaw Codelab to stand up a persistent agent, then keep the create-instance quickstart handy for the stop, log, and delete path.

OpenClaw, from laptop to a public URL

The CLI lifecycle, without the agent

Preview caveats

Sources