Short answer: Hermes Agent is an open-source AI agent from Nous Research that runs on your own machine or server and talks to you through apps like Telegram, Slack, WhatsApp or the command line. Its defining feature is a built-in learning loop: it turns finished tasks into reusable skills, improves them as it uses them and keeps a searchable memory of past sessions. It is MIT-licensed and works with almost any model, from Claude and GPT to open models on a local endpoint (Hermes Agent docs).
Hermes Agent belongs to the fastest-growing category of AI tools in 2026: personal agents that stay on, live inside your messaging apps and act on your behalf. The GitHub repository passed 250,000 stars by early October 2026, and its last release before this guide, v0.21.5, shipped on September 24. If you are mapping the wider landscape of assistants and agents, start with our AI tools hub. This guide covers what Hermes Agent does, how its learning loop and memory work, how to give it a computer safely and how it compares with OpenClaw.
What is Hermes Agent?
Nous Research describes Hermes Agent as "the self-improving AI agent", with a closed learning loop that creates skills from experience, improves them during use and builds a deepening model of who you are across sessions. In practice it is three things at once:
- An agent runtime that plans, calls tools, runs commands and remembers.
- A gateway that connects one agent to 20+ messaging platforms, so you can message it from wherever you already work.
- A skills system that grows over time, either from the community Skills Hub or from the agent itself.
Unlike cloud agents such as Grok Bot, Hermes Agent is self-hosted. You decide where it runs, which model it uses and what it can touch.
How does the learning loop work?
Most agents start every task from zero. Hermes Agent is built to get cheaper and better at the work you repeat. According to the documentation, the loop has four parts:
- Autonomous skill creation. After a complex task, the agent can extract the procedure into a reusable skill.
- Skill self-improvement. When it uses a skill again, it can refine it based on what happened.
- Agent-curated memory with periodic nudges. The agent decides what is worth keeping and is prompted to persist knowledge.
- Cross-session recall. Past sessions are searchable with SQLite full-text search (FTS5) and summarized by the model, and user modeling builds a picture of your preferences over time.
The result is closer to a colleague who learns your processes than to a chatbot. Skills are portable too: Hermes Agent is compatible with the agentskills.io format, so skills can be shared through the Skills Hub. For a deeper look at why memory changes what agents can do, see our guide on how AI agent memory works.
Which apps and models does it work with?
One gateway connects the agent to more than 20 platforms. The documentation lists the CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, email, SMS, Microsoft Teams, Google Chat and Home Assistant, plus several platforms popular in China such as DingTalk, Feishu, WeCom and Weixin.
On the model side, Hermes Agent is not tied to one provider. It works with Nous Portal, which gives access to more than 300 models, OpenRouter, OpenAI or any compatible endpoint, including open models running locally. Switching models is a configuration change, not a migration.
How does Hermes Agent use a computer?
Hermes Agent can control a desktop the way a person does, clicking, typing, dragging and scrolling, on macOS, Windows and Linux. The interesting part is that it works in the background. The computer use documentation says your cursor doesn't move, keyboard focus doesn't change and your virtual desktops don't switch while it works.
Under the hood, Hermes speaks MCP to cua-driver, an open-source background computer-use driver, which uses each operating system's accessibility stack. And unlike most computer-use integrations, it works with any tool-capable model rather than one vendor's computer-use API.
There are guardrails. Destructive actions need approval through the same gate as dangerous shell commands, some key combinations are blocked outright, such as emptying the trash or locking the screen, and the agent is instructed never to click permission dialogs or type passwords.
Where should the agent run? The 7 terminal backends
Computer use covers the screen. For commands, Hermes Agent offers seven terminal backends, and the choice decides how much of your system the agent can reach:
| Backend | What it is | Isolation |
|---|---|---|
| local | Runs commands on your machine as your user | None |
| Docker | Each session runs in containers | Strong |
| SSH | Runs on another machine or VM | Depends on the host |
| Singularity | HPC-style containers | Strong |
| Modal | Managed cloud sandbox | Strong |
| Daytona | Managed cloud sandbox | Strong |
| Vercel Sandbox | Managed cloud sandbox | Strong |
The local backend is convenient for a first test and risky for anything else, because the agent acts with your permissions on your files. For real work, put it in a container or a sandbox, or give it its own machine over SSH. NVIDIA's NemoClaw blueprints also list Hermes Agents among the frameworks they can wrap with policy controls, which shows where enterprise deployments are heading.
Hermes Agent vs OpenClaw: which one should you pick?
OpenClaw is the other big open-source personal agent, and people compare the two constantly. Both are open source, MIT-licensed and self-hosted, and both connect to your messaging apps. The differences are in philosophy:
| Hermes Agent | OpenClaw | |
|---|---|---|
| Maker | Nous Research | Community project created by Peter Steinberger, now run by a foundation |
| Core idea | An agent that learns: self-created and self-improving skills | A gateway and runtime: you shape the agent with workspace files |
| Memory | Agent-curated memory and full-text search across sessions | Markdown files you can read and edit, such as MEMORY.md and daily logs |
| Skills | Self-generated plus the agentskills.io Skills Hub | ClawHub skills and plugins |
| Isolation options | 7 terminal backends, from local to cloud sandboxes | Loopback-only gateway by default, sandboxing you configure |
Pick Hermes Agent if you want an assistant that improves at your recurring work and you value model freedom. Pick OpenClaw if you want the larger ecosystem and full control of the agent's instructions in plain files. Many developers try both. Whichever you choose, the security rules below apply.
How do you install Hermes Agent?
The documented path is short:
- On Linux, macOS or WSL2, run the install script from the official site (
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash). Windows has its own PowerShell command, and there is a Hermes Desktop installer. - Run
hermes setup --portalfor the fastest setup through Nous Portal, or configure your own provider. - Connect one messaging channel first, for example Telegram, and test with low-risk tasks.
- Move the terminal backend to Docker or a cloud sandbox before you give it anything that matters.
What should you check before trusting it with real work?
A personal agent is a command executor attached to an inbox, so treat it like one:
- Isolate it. Use Docker, a cloud sandbox or a separate VM, never the local backend for sensitive work.
- Limit who can talk to it. Only allow your own accounts on each channel.
- Use dedicated accounts and scoped keys. Give the agent its own email, phone number and read-only API keys where possible.
- Keep approvals on. Let destructive actions and shell commands wait for your OK until you trust the setup.
- Review its skills. Self-generated skills are code-like procedures. Read what it saves, as you would review a pull request.
The skill behind a useful agent
Hermes Agent rewards people who can describe a workflow precisely, choose the right isolation and review what an agent writes. Those are the core skills of AI engineering, and they can be learned. If you want to build them with mentors and real projects, compare our AI programs and pick the one that fits your starting point.