Hermes Agent on GCP - An Agent That's Actually Yours, Billed the Way GCP Already Bills You
If your team already lives in Google Cloud, spinning up a second cloud account just to run an AI agent is the kind of extra step that never quite gets done. It sits on a to-do list next to "clean up the IAM roles" for six months.
Meetrix put Hermes Agent on the GCP Marketplace so that isn't necessary. It's a single Compute Engine deployment inside your existing GCP project, billed the way GCP already bills you: a combined estimate for the licence fee and the VM before you deploy, not a separate invoice from a separate vendor. Hermes Agent ships pre-installed and running, with a 5-day free trial worth up to USD 50 in licence credit to try it before anything's charged for real. Ready to look? Open the Hermes Agent listing on GCP Marketplace.
What Hermes Agent Actually Does
Hermes Agent is a self-improving agent built by Nous Research, and that's a specific claim, not marketing copy. Ask it to do something new a few times and it writes itself a reusable skill for next time, instead of re-deriving the same steps from zero on every run.
It also doesn't forget you the moment a session ends, which is the thing most agent tools quietly get wrong. Memory persists across sessions through SQLite's FTS5 full-text search combined with LLM summarization, so a decision from three weeks ago stays retrievable rather than vanishing when you close the chat window.
One gateway process serves Telegram, Discord, Slack, WhatsApp, and Signal at once, plus a terminal TUI if you'd rather talk to it directly over SSH. That's one memory, one config, five surfaces, not five separate bots you're keeping in sync by hand. A built-in cron scheduler handles recurring jobs, and parallel subagents split larger tasks into pieces that run at the same time instead of queuing behind each other. And it isn't locked to one model vendor: run hermes model and switch providers whenever you want, Google Vertex AI included. Which model you route through shapes the running cost as much as the machine type does, so if you haven't settled on one yet, our rundown of the best open source LLMs for self-hosting is a reasonable place to start.
Why Vertex AI Is Worth Considering Here
How Deployment Works
Doing this yourself means provisioning a Compute Engine instance, installing Python 3.11 (most base images still ship 3.10), Node.js, ripgrep for skill search, ffmpeg for voice messages, a systemd unit so the gateway survives a reboot, and certbot for SSL, before you've sent a single message. Through the Marketplace listing, or the full walkthrough with screenshots in our GCP Developer Guide, it's four steps:
- Find and Launch From GCP Marketplace Open the Hermes Agent listing, select your GCP project, accept the terms, and optionally activate the 5-day free trial before deploying.
- Configure the Deployment Form Pick a zone close to your users, choose a machine type (General purpose is preselected and fine to start with), and set your domain, admin email, username, and password under Application Settings.
- Point DNS at the External IP Copy the external IP from the VM instance's network interface and add it as an A record with your DNS provider. SSL issues itself automatically once that propagates.
- Connect a Model Provider and Log In
SSH in, run
sudo docker exec -it hermes hermes model, and pick a provider. Then sign in to the dashboard with the admin credentials you set and start pairing channels.
What Meetrix Brings to This Deployment
- No Dependency Wrangling - Python 3.11, Node, ripgrep, and ffmpeg all have to be installed and version-matched before Hermes Agent will even start. This image ships with all of it already running.
- SSL That Handles Itself, With a Fallback - Automatic Let's Encrypt issuance once DNS propagates, and
/root/certificate_generate_standalone.shon the instance if it doesn't catch on the first try. - Credentials You Set, Not a Shared Default - Admin username and password are stack parameters you choose at deploy time, not something to change later because it started on a default.
- A Combined Cost Estimate Before You Commit - GCP shows the licence fee and infrastructure cost together on the deployment form, before you click Deploy, plus the 5-day trial to test it against real usage first.
- People Who've Actually Run This Image - Support goes to support@meetrix.io with a response inside 12 hours, from engineers who deploy this on GCP regularly.
Who Is Hermes Agent on GCP Right For?
This fits a fairly specific situation better than a general one: you want an agent you actually own, and Google Cloud is already where your team's infrastructure and billing live. More concretely:
- A team already running its VPC, IAM, and billing through GCP that doesn't want a second cloud account just for an agent
- Anyone planning to use Vertex AI as the model provider, where hosting and inference end up on one bill instead of two
- A solo founder who wants a Slack assistant that remembers last week's thread, not one that starts cold every morning
- An agency running client bots in Discord, where parallel subagents split work across accounts instead of paying for a deployment per client
- A developer with API keys already sitting around for a provider Hermes supports, who'd rather not hand them to a third-party SaaS
- A compliance-conscious org that needs conversation history and API keys to stay inside a GCP project it already governs
Hermes Agent on GCP by Meetrix vs Alternatives
If you're weighing agent platforms against each other rather than deciding how to host this one, our comparison of Flowise, Langflow and n8n covers the visual-builder end of that choice. The table below is narrower on purpose: one product, three ways of running it.
| Feature | Hermes Agent on GCP by Meetrix | Hosted AI Agent SaaS | Self-Managed Install |
|---|---|---|---|
| License | MIT, fully open source | Proprietary | Open source, but you own every step of running it |
| Pricing Model | Combined GCP + licence fee estimate shown before deploy, plus your own model cost, no markup | Per-seat or usage markup layered on top of model cost | Server cost plus your own model cost |
| Free Trial | 5 days, up to USD 50 in licence credit | Varies, often capped by message count | None, you're paying for compute from day one |
| Model Provider Fit | Native Vertex AI support, plus OpenAI, Anthropic, OpenRouter, Bedrock | Usually locked to the vendor's own model | Whatever you wire up yourself |
| Platforms From One Deployment | Telegram, Discord, Slack, WhatsApp, Signal, and a terminal TUI, one gateway | Usually one, more as paid add-ons | Whatever you wire up yourself |
| Data Location | Your own GCP project | Vendor's infrastructure | Your server |
| Support | Meetrix engineers, response inside 12 hours | Support tiers, often gated by plan | Community forums only |
Resources
Running on AWS instead? The Hermes Agent AWS Marketplace listing covers the same product as a CloudFormation stack, if that's the cloud your team already uses.
Under the Hood
Worth knowing what's actually running before you point it at a model API key.
Docker Container
Hermes Agent runs inside a container named hermes on the instance. sudo docker exec -it hermes hermes model reaches the CLI to set a provider, and the same container serves the web dashboard.
nginx in Front
nginx terminates SSL and enforces HTTP Basic Auth before a request reaches the container, using the admin credentials you set in Application Settings (stored in /etc/nginx/.htpasswd, changeable afterward over SSH).
A Persistent Disk, Not Ephemeral
Memory, sessions, and skills live on the instance's boot disk. Stop the VM to save on compute overnight and the data is still there when it comes back. If the disk fills, resize it in the console rather than deleting anything.
FTS5, Not Vector Search
Conversation history persists through SQLite's FTS5 full-text search combined with LLM summarization, stored on the instance itself, not shipped to a third-party vector database.
An Ephemeral External IP by Default
The deployment form leaves the external IP as Ephemeral by default, which is fine since your DNS A record points at whatever IP the instance has. Reserve a static IP instead if you'll be stopping and restarting the VM often.
SSL Renews Itself
Certificates come from Let's Encrypt automatically once DNS resolves. If it isn't ready in time, /root/certificate_generate_standalone.sh generates one manually.
One Thing Hermes Agent Won't Do For You
Memory recall here is lexical, not semantic. FTS5 matches the words you actually used, not the meaning behind them. Ask about "the pricing outage" three weeks after discussing it as "the billing incident," and there's a real chance it doesn't come up, because the two phrases don't share enough words for FTS5 to connect them. Vector-embedding memory handles that kind of paraphrase better. Hermes Agent's approach is cheaper and faster to run, and most people describe the same thing the same way twice anyway, but it's worth knowing before you hit it mid-incident rather than during one.
Frequently Asked Questions
What is Hermes Agent used for on GCP specifically?
The same things it's used for anywhere: a Slack or Discord assistant that remembers last week's conversation, a Telegram bot for personal automation, scheduled jobs through the built-in cron scheduler. The GCP angle matters if your team already runs its infrastructure on Google Cloud and would rather not open a second cloud account just to host an agent.
Do I need my own model API key?
Yes. The image doesn't bundle model access. Run hermes model after deployment and point it at whichever provider you use, Nous Portal, OpenAI, OpenRouter, Anthropic, Google Vertex AI, or AWS Bedrock. You're billed by that provider directly, not by Meetrix.
What does this actually cost on GCP?
Two line items: the Meetrix licence fee and the standard GCP charge for the VM and disk you select. Google shows both combined as an estimated monthly cost on the deployment form before you click Deploy, so there's no guessing. A 5-day free trial with up to USD 50 in licence fee credit is available, though the underlying VM and disk charges still apply during the trial.
Which machine type should I pick?
The default General purpose machine type preselected on the deployment form is enough to get started and get a feel for how the agent behaves. Size up if you're running several concurrent automations or connecting multiple messaging channels with real traffic. If you hit a CPU quota error at deploy time, that's a regional Compute Engine quota, not a Hermes Agent limit, and it's fixed by requesting a quota increase or picking a different zone.
Does it remember conversations between sessions?
Yes. Memory persists using SQLite's FTS5 full-text search combined with LLM summarization, stored on the instance's own persistent disk. It survives a VM stop and restart, which matters if you're pausing the deployment to save on compute overnight.
Can I use Vertex AI as the model provider since I'm already on GCP?
Yes, and it's arguably the more natural setup if your billing and IAM are already centered on Google Cloud. Vertex AI is one of the providers hermes model lists directly, alongside OpenAI, OpenRouter, Anthropic, and AWS Bedrock, so you're not forced into a second vendor relationship just to get the agent talking.
Where does my data live?
On the Compute Engine instance inside your own GCP project, not on Meetrix's infrastructure. Conversation history, the memory database, and any model API keys you configure stay on that instance under its own file permissions. Meetrix doesn't have access to the running deployment once it's launched.
What if the automatic SSL setup fails?
The instance tries to issue a Let's Encrypt certificate for the domain you provide during deployment, which only works once DNS has actually propagated to the instance's external IP. If it fails, or DNS wasn't ready in time, SSH into the instance and run sudo bash /root/certificate_generate_standalone.sh to issue it manually.
Run Your Own Agent on GCP
Stop paying per seat for an agent you don't control. Deploy Hermes Agent on Google Cloud in minutes, billed alongside the infrastructure you already run.
Deploy on GCP Marketplace