Evan Lin for Google Developer Experts

Posted on May 2 • Originally published at evanlin.com on May 2

GCP in Action: Building a Persistent AI Assistant with GCE, Hermes Agent, and Telegram

#agents #ai #googlecloud #tutorial

Background

After solving the LINE Bot's Vertex AI migration, I started thinking: Could there be an AI assistant that is "more proactive" and "has long-term memory"? At this time, I set my sights on NousResearch's open-source Hermes Agent.

Unlike a typical Chatbot, Hermes is designed as an "operating system that breathes". It can execute Shell commands, write Python scripts, manage long-term memory, and even stay in touch with you via different Gateways (Telegram, Discord) at any time.

To make it available 24/7, I chose to deploy it on Google Compute Engine (GCE). This article will document the deployment process from scratch, as well as the pitfalls I encountered when configuring the latest Gemini 2.5 Flash model.

Environment Parameter Preparation

Before you start, please make sure you have these necessary parameters:

PROJECT_ID: YOUR_PROJECT_ID
LOCATION: global
GOOGLE_API_KEY: YOUR_GOOGLE_API_KEY (Obtained from Google AI Studio)

Step 1: Create a GCE Instance

Hermes Agent needs some computing power to handle tool use. It is recommended to use the e2-medium specification.

gcloud compute instances create hermes-agent-vm \
    --project=YOUR_PROJECT_ID \
    --zone=us-central1-a \
    --machine-type=e2-medium \
    --image-family=ubuntu-2204-lts \
    --image-project=ubuntu-os-cloud \
    --boot-disk-size=30GB \
    --metadata=startup-script='#!/bin/bash
        apt-get update
        apt-get install -y git curl python3-pip python3-venv nodejs npm
    '

Step 2: Install Hermes Agent

After SSHing into the VM, use the official one-click installation script directly.

Enter the VM:

gcloud compute ssh hermes-agent-vm --zone=us-central1-a

Execute the installation:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.bashrc

Step 3: Configure Gemini 2.5 Flash (SOP Practice)

This is the most likely place to step on a landmine in the entire exercise. Hermes may default to pointing to non-existent or outdated model identifiers.

Create a configuration file: In ~/.hermes/config.yaml, we must precisely specify Gemini 2.5 Flash, and do not include the google/ prefix.
Set the API Key: Write the key and permission settings in ~/.hermes/.env:

Step 4: Connect to Telegram and Background Persistence

To prevent the Agent from disappearing after the SSH connection is lost, we use Systemd to manage it.

Create a Systemd service (/etc/systemd/system/hermes.service):

[Unit]
Description=Hermes Agent Gateway
After=network.target

[Service]
Type=simple
User=root
Environment=HOME=/root
Environment=PYTHONUNBUFFERED=1
ExecStart=/usr/local/lib/hermes-agent/venv/bin/hermes gateway run
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

Start the service:

sudo systemctl daemon-reload
sudo systemctl enable hermes
sudo systemctl restart hermes

Blood and Tears in the Migration Process: Why Isn't My Agent Responding?

Even with the correct configuration, I still encountered the dilemma of "the Agent reads messages but doesn't reply". After checking the logs (journalctl -u hermes), I found several deep pitfalls:

Pitfall 1: The 404 Ghost of Gemini 3.0

I tried to pursue the latest version when configuring, and used gemini-3-flash-preview. As a result, the logs spewed out a bunch of 404 Model Not Found. Reason: The internal auxiliary_client.py of Hermes hardcodes many gemini-3-flash-preview as the default value. When these auxiliary functions (such as generating titles) report errors, it will affect the reply logic of the entire Gateway. Solution: Manually define all auxiliary models as gemini-2.5-flash in config.yaml, or directly patch the source code with sed.

Pitfall 2: Prefix Confusion of Model Identifiers

In different SDKs, some people use google/gemini-2.5-flash, and some people use gemini-2.5-flash. Experience: In Hermes' Gemini Provider, using the short name gemini-2.5-flash directly is the safest. Adding google/ will instead cause API routing errors.

Pitfall 3: Conflict between Systemd and "Processes Already Running"

When you manually run hermes gateway and then start the service, the system will report Gateway already running (PID xxxx). Solution: Before ExecStart in Systemd, you can add an ExecStartPre=/usr/bin/pkill -9 -f hermes || true to ensure a clean environment every time you start.

Summary

Now, my dedicated Hermes Agent is running stably on GCE and is available via Telegram at any time. It can not only help me find information, but also run some simple computing scripts for me directly on the cloud VM.

This deployment taught me: In the face of rapidly updating models, the official documentation (or MCP tool query) is the only truth. Don't blindly pursue the latest version number; ensuring that the identifier matches the current API environment is the key to stable operation.

If you also want a 24-hour AI digital double, get a machine set up according to this SOP!

DEV Community