Hermes Agent - Complete Local Setup with Gemma 4

Published: April 16, 2026 | Tags: hermes-agent, gemma-4, llama.cpp, local-ai, tutorial, docker

Table of Contents

What is Hermes Agent?
Prerequisites
Installation Steps
Configuration
Custom Persona
Key Features
Resources

What is Hermes Agent?

Hermes Agent (by NousResearch) is a self-improving AI agent that:

Learns from experience and builds skills over time
Remembers users across sessions
Runs anywhere locally - VPS, Mac, PC
Works with 200+ models
Integrates with Telegram, Discord, Slack, Signal, Email

Described as "the best alternative to OpenClaw" - it's an open-source agentic harness similar to OpenAI's Codex.

Prerequisites

Docker installed
LM Studio or llama.cpp server running with Gemma 4
16GB+ RAM recommended
Tavily API key (optional, for web search - 1K free requests/month)

Installation Steps

Step 1: Create Hermes Folder

mkdir ~/.hermes

Step 2: Start Docker with LM Studio Access

docker run -it -v ~/.hermes:/tmp/data \
  -p 3000:3000 \
  -v /tmp/data:/tmp/data \
  --add-host=host.docker.internal:host-gateway \
  nousresearch/hermes-agent:latest

Note: The --add-host=host.docker.internal:host-gateway flag allows the container to access your host machine's LM Studio server.

Step 3: Run Setup

hermes-agent setup

Configuration

Provider Setup

Choose Custom Endpoint → points to LM Studio
Host: host.docker.internal
Port: 8080 (or your LM Studio port)
Version: v1
API Key: (none needed for local)

Model Settings

Setting	Recommended Value	Notes
Context Length	Auto-inferred	Automatically detected from model config
Tool Calling Max Iterations	20	Prevents infinite loops
Tool Compression Threshold	50%	Compresses when context is 50% full
Session Reset Modes	Default	Configurable per session

Select Tools

✅ Web Search - Use Tavily API (1K free requests/month)
✅ Web Scraping - Built-in capability
✅ Browser - Headless Chromium included
✅ Skills - Built-in + community skills

Custom Persona

Create your agent's personality by editing ~/.hermes/prompts/soMD:

Keep your responses to the point, clear, eliminate fluff.
Keep it ranked, your audience is ML/AI engineers.
Be friendly and responsive.

Key Features

Feature	Description
Memory	Automatic recall from previous sessions
Skills	Self-creates skills during conversations
Self-Improving	Gets better over time based on interactions
Web Search	Uses Tavily API for real-time information
Web Scraping	Headless Chromium for page extraction
Multi-Platform	Telegram, Discord, Slack, Signal, Email
Dashboard	Web UI to monitor agent status and sessions

Video Result: The presenter tested Gemma 4 (6B params, quantized 4-bit) on an M4 MacBook Pro with 48GB unified memory. The agent correctly: - Did web research on Bulgarian hip hop artists - Retrieved technical details on MiniMax 2.7 model with citations - Analyzed recent AI papers (SKU-code framework) - Created custom skills during conversations

Starting the Agent

Start without Setup (already configured)

docker run -it -v ~/.hermes:/tmp/data \
  -p 3000:3000 \
  -v /tmp/data:/tmp/data \
  --add-host=host.docker.internal:host-gateway \
  nousresearch/hermes-agent:latest

Access Dashboard

Navigate to http://localhost:3000 to see:

Current model and status
Active sessions and message history
Search API usage and citations
Skills created during conversations

Resources

Hermes Agent GitHub
LM Studio - Run Gemma 4 locally
Gemma 4 Model
Tavily API - Web search (1K free/month)
OpenClaw Discord

Based on tutorial by Venelin Valkov (@venelin_valkov). Video embedded above.