What Is Hermes Agent? The Claude-Compatible Runtime Explained for Claude Code Users
Hermes Agent is an open-source self-improving AI runtime from Nous Research. Here's what Claude Code users need to know about it in 5 minutes.
If you spend your days inside Claude Code, there is a good chance you have hit the same wall twice this week: the session ended, and so did the agent's memory of everything it just learned. You opened a new terminal, watched it re-read your CLAUDE.md, and silently grieved the context it lost. Hermes Agent, an open-source runtime from Nous Research, was built to sit next to Claude Code and fix exactly that problem.
This article is the entry point to our Hermes foundations series. It answers the question Claude Code users keep asking us: what actually is Hermes, who made it, and why should I care when my claude CLI already works fine?
Key Takeaways
- Hermes Agent is an open-source, self-improving agent runtime from Nous Research with 113,000 GitHub stars and MIT license.
- It is model-agnostic but ships with first-class Claude support, including the
anthropic>=0.39.0SDK and direct reads from the Claude Code credential store. - Hermes is a persistent server, not a CLI session. It runs on a VPS, accumulates markdown memory, and keeps working when your laptop closes.
- Skills use the same
agentskills.iostandard as Claude Code, so well-formed SKILL.md files are format-compatible between the two (minor tagging tweaks may be needed). - Hermes can spawn Claude Code as a subagent via print mode or tmux, making it a scheduler and supervisor for your Claude sessions.
- The marketing page is get-hermes.ai; docs live at hermes-agent.nousresearch.com/docs.
- Self-hosting typically runs $4 to $8 per month on a small VPS; a hosted SaaS option called Agent37 starts at $3.99 per month.
What Hermes Actually Is
Hermes Agent is a Python runtime, published to PyPI as hermes-agent (version 0.10.0 at time of writing), that hosts autonomous LLM agents. It is not a model family. The name overlaps with Nous Research's Hermes fine-tunes but the runtime is a separate product: a server-side framework that talks to whatever frontier model you point it at.
Under the hood, the repository is roughly 87.5% Python and 8.5% TypeScript, requires Python 3.11 or newer, and installs with a single shell command:
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
After installation, three entry points become available: hermes, hermes-agent, and hermes-acp. The first is the day-to-day CLI. The third is an adapter for the Agent Communication Protocol, which lets Hermes plug into VS Code, Zed, and JetBrains IDEs.
The Defining Feature: A Learning Loop
Most agent frameworks treat skills as static files you write once and commit to a repo. Hermes treats them as living knowledge. When the agent completes a tricky task, it can write a new SKILL.md to its own skills directory, capturing what worked. Over weeks of use, your Hermes instance accumulates a library of procedures it wrote for itself.
Combined with a markdown memory store indexed by SQLite FTS5, this is what Nous Research means by "the agent that grows with you." We go deeper on the memory side in the Hermes memory deep dive.
Who Made It and Why
Nous Research, the lab behind the Hermes and DeepHermes fine-tunes, open-sourced the runtime in 2025 under the MIT license. The project has grown quickly: 113,000 stars, 16,400 forks, 615 contributors. The positioning is explicit in the tagline on the marketing site: "the self-improving AI agent that runs on your server." The emphasis on your server matters. Hermes was not designed as a SaaS-first product. Self-hosting is the first-class path; the hosted Agent37 option exists for people who do not want to run a VPS.
How It Relates to Claude Code
This is where most of our readers land. The short answer: Hermes is not a replacement for Claude Code. It is a complementary layer.
Claude Code is a local CLI that lives in your terminal for the duration of a coding session. Hermes is a long-running daemon on a server. They integrate in three important ways.
1. Hermes Uses Claude as a First-Class Provider
The default configuration looks like this:
model:
provider: "anthropic"
default: "claude-sonnet-4-6"
You can launch a chat with Claude Sonnet 4.6 directly:
export ANTHROPIC_API_KEY=***
hermes chat --provider anthropic --model claude-sonnet-4-6
The aliases claude and claude-code both resolve to the same underlying provider path, which is a convenience detail Claude Code users appreciate. More on this in running Hermes on Claude Sonnet 4.6.
2. It Reads the Claude Code Credential Store
If you already have Claude Code installed and logged in on your server, Hermes can read the existing credential store directly. No token copying, no duplicate OAuth flows. Your Claude Pro or Max plan simply carries over.
3. It Can Delegate to Claude Code as a Subagent
Hermes ships a bundled skill at skills/autonomous-ai-agents/claude-code/SKILL.md that teaches the agent how to spawn claude -p (print mode) or drive a tmux-based PTY session. The Hermes orchestrator decides what to do, the Claude Code subagent executes the code task, and the results flow back into Hermes memory. We cover this workflow in depth in the comparison pieces later in the series.
The Skills Story
Hermes skills use the agentskills.io open standard. Frontmatter looks exactly like a Claude Code skill:
---
name: test-driven-development
description: Use when implementing any feature or bugfix, before writing implementation code. Enforces RED-GREEN-REFACTOR cycle with test-first approach.
version: 1.1.0
author: Hermes Agent (adapted from obra/superpowers)
license: MIT
metadata:
hermes:
tags: [testing, tdd, development, quality, red-green-refactor]
related_skills: [systematic-debugging, writing-plans, subagent-driven-development]
---
Because both tools follow the same open spec, a well-formed SKILL.md is format-compatible between them. In practice you may need to adjust the metadata.hermes.tags block versus a Claude Code equivalent, but the structure and body travel cleanly. We walk through the spec in writing your first Hermes SKILL.md.
Architecture Highlights
A few facts worth knowing before you install:
- Tools: 47 built-in tools plus native MCP support, with the same MCP servers you use in Claude Code.
- Execution backends: Six options for running terminal commands, including local, Docker, SSH, Daytona, Singularity, and Modal.
- Messaging gateways: Eight, including Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Feishu, and DingTalk.
- Voice mode: Push-to-talk in 20 languages, useful when the laptop is closed.
- Scheduling: A built-in cron scheduler that accepts natural-language task definitions.
The footprint is deliberately modest. Hermes runs comfortably on a $4 to $8 VPS, which we cover step-by-step in installing Hermes Agent on a $5 VPS.
When to Pick Up Hermes
Reach for Hermes when you want your Claude workflow to keep running after you log off. It is the right tool if you are scheduling daily reports, watching inboxes, coordinating multiple parallel Claude sessions, or building an agent that genuinely remembers last Tuesday. It is the wrong tool if you just need a coding copilot in your editor; stay in Claude Code for that.
Most of our readers end up using both.