You open a fresh chat with your AI assistant. And you explain everything again. The project name, the repo layout, the way you name your branches. You said all of it yesterday, in a conversation the tool has forgotten. Hermes Agent goes straight at that problem.
Hermes Agent is an open source framework for autonomous AI agents, released by Nous Research in February 2026 under the MIT license. It does not run as a chatbot in a browser tab. It runs as a service installed on your own machine or server, always on. It keeps its memory between sessions, carries out multi-step tasks, and writes its own method sheets that it reuses later.
You will see how its work loop turns, what its memory really does, what those skills look like, what it costs to run, and where it disappoints. No prior knowledge of AI agents needed.
What Hermes Agent actually is
It is an agent, not a conversational assistant. The difference is initiative. A chatbot answers a question and stops. An agent takes a goal, breaks it apart, runs actions on your machine, checks the result, and tries again when it fails.
Hermes Agent runs as a background process, usually on a virtual private server (a computer you rent from a host, powered on around the clock). You send it instructions from Telegram, Discord, Slack or a terminal. It keeps working after you close the conversation.
Nous Research also publishes a family of language models called Hermes. Hermes Agent is not a model. It is the shell that puts a model to work, theirs or one you pick yourself, by adding memory, tools and access to your machine.
The project ships under the MIT license, the most permissive of the free software licenses. You can read the code, change it, use it commercially. Everything sits in the Hermes Agent GitHub repository.
How it works, step by step
Once installed, Hermes Agent creates its working folder at ~/.hermes/ and starts a loop that never really ends. It runs in five beats.
- Input. A request arrives through the command line, a connected messaging app, or a scheduled job.
- Planning. The agent sends the request to the configured language model, along with its current memory and the list of available tools. The model picks the next action.
- Execution. The agent uses the chosen tool: terminal, file editor, web browser. The result goes back into the loop.
- Memory. Results and reasoning are written to a local database, sometimes into structured memory files.
- Iteration. The cycle repeats until the task is done. Then the agent waits for the next trigger, without shutting down.
That persistence changes the kind of work you can hand over. A job spread across several days does not restart from scratch every morning.
Memory in four layers
This is the most carefully built part of the product, and the reason behind the Nous Research tagline. Hermes Agent does not store everything in one place, because not all memories have the same shelf life.
- Two short files,
MEMORY.mdandUSER.md, holding your environment and your preferences. They are injected at the start of every session. - A session archive in a SQLite database (
~/.hermes/state.db), indexed with FTS5, the full text search engine built into SQLite. Every past conversation stays searchable. - A skills folder, which holds the know-how: how to do a given thing, in what order.
- External memory providers such as Honcho, Mem0, OpenViking or Supermemory, which sharpen the user profile over time.
In practice, you stop re-explaining your context. The agent knows where your code lives, how you deploy, which commit message format you use.
The skills it writes on its own
After a hard task, the kind that takes five tool calls or more, Hermes Agent pauses and writes a structured document describing what it just did. The project calls that a skill.
Take a deployment that fails. The agent pokes around, finds the cause, fixes it. Then it records the procedure. Next time a similar problem shows up, it opens its own sheet instead of starting the investigation over.
These sheets are plain Markdown files stored in ~/.hermes/skills/. They follow the open agentskills.io format, which makes them portable to other compatible agents. You can also install community-written skills from the Skills Hub, with commands like /github-pr-workflow.
The agent updates its own skills when it decides they are stale. If you edit a sheet by hand, expect it to be rewritten on the next improvement cycle. Keep a copy outside ~/.hermes/ for the sheets you care about.
Tools, models and hosting
Hermes Agent ships with more than 40 built-in tools. Web search, terminal commands, reading and patching files, browser control, image analysis, image generation, text to speech, delegation to sub-agents. Tools register themselves when they load.
You can add your own by dropping an extension into the expected folders. It is Python code, a language beginners can pick up, which puts extending the product within reach of someone who is not a systems engineer.
On the brain side, any endpoint compatible with the OpenAI API works. Nous Portal, OpenRouter, Anthropic, or a local Ollama instance if you want nothing to leave your building. The Model Context Protocol, the standard for plugging external tools into an agent, is supported over stdio or HTTP and configured in a single config.yaml file.
Hermes Agent requires a model that can handle at least 64,000 tokens of context. A token is a fragment of a word, the unit a model chops text into. That requirement applies to the model you plug in, not to your server: a small VPS is fine when the model runs elsewhere.
For running commands, you pick from six backends: local for speed, Docker for isolation, SSH for a remote server, plus Daytona, Singularity and Modal. On a server exposed to the internet, Docker is the sane default, because the container stands between the agent and the rest of the machine.
The software itself costs nothing. Your spending comes from the server and from model calls, and it climbs fast if you leave browser automation or several parallel sub-agents running.
Hermes Agent next to the other tools
The comparison people reach for first is OpenClaw, another self-hosted autonomous agent. The two look alike from a distance, but their logic differs.
| Criterion | Hermes Agent | OpenClaw |
|---|---|---|
| Main goal | Improve itself by creating its own skills | Run tasks stated in plain language, directly |
| Memory | Four layers: prompt files, SQLite archive, skills, external providers | Long-running conversational memory |
| Persistence | Continuous loop with scheduled jobs and cron | Always-on assistant |
| Hosting | Local machine, VPS, serverless backends such as Modal or Daytona | Local machine, Raspberry Pi or VPS |
If you already run OpenClaw, the hermes claw migrate command imports your settings, memories, skills and API keys during the initial setup. You do not reconfigure your model providers or your messaging connections.
Paperclip sits one floor up. It is an orchestration layer that coordinates several agents with defined roles and a hierarchy. Hermes Agent can work inside it as a worker. The two complement each other more than they compete.
One last distinction: do not mix this family of tools up with coding assistants. Claude Code, Anthropic's command-line programming assistant, works alongside you on a project, under your eyes. Hermes Agent works while you do something else, and its scope goes well past code.
What people actually use it for
Four uses stand out, all resting on the same mechanic: the agent holds context over time and chains steps without you watching the screen.
- Development automation: reviewing pull requests, running tests, refactors that stretch over several days.
- Research and data analysis: it browses, runs code, stores what it finds and comes back to it later.
- Scheduled work: a weekly report, an overnight watch on a topic, a morning digest pushed to Telegram or Discord.
- A persistent personal assistant: it learns your preferences and takes over repetitive chores.
What links those four is tempo. They are jobs that stretch out, not one-off questions. For a one-off question, a regular chatbot is still quicker to reach for.
Limits worth knowing before you start
Hermes Agent picks control over convenience. That choice has a price, and you want to know it before installing rather than after.
You need to be able to run a machine. Keeping an agent alive around the clock means understanding Linux, system services and containers, at least enough to diagnose a failure on a Sunday night.
An autonomous agent with access to your terminal and your API keys widens the attack surface a lot. Hermes Agent blocks certain dangerous commands and asks for confirmation on sudo, but those guardrails can be bypassed in some container setups. Hardening the host, isolating credentials and watching the logs is your job.
The bill can climb without warning. A light setup fits on a small VPS, but browser automation and parallel sub-agents eat CPU, memory and tokens. At scale, the model becomes the main line item, not the hosting.
The agent is never better than its model. Made-up facts, misread tool output, a thread of logic lost halfway through a long task: none of that goes away because you bolted memory onto it.
Two more down-to-earth constraints. The learning loop fires after complex tasks, so it gives you nothing if your requests never resemble each other. And there is no native Windows build: you go through WSL2. On Android, Termux works, with voice and media features cut back.
How to get started
First decision: where the agent will live. An install on your laptop dies when you close the lid, which cancels the whole point of persistence. A VPS is the coherent choice.
You need a Linux server, a Mac or a WSL2 instance, with Python 3.11. Docker is optional, but it is the isolation layer that spares you nasty surprises. The official installer handles the rest.
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bashThat command installs the dependencies (uv, Python, Node.js, ripgrep, ffmpeg) and creates the ~/.hermes/ folder. You then pick your model provider with hermes model, connect a messaging app with gateway setup, and switch on an external memory provider with hermes memory setup if you want one.
Last step, and the one that matters most for continuity: run the agent as a system service so it comes back up on its own after a reboot. If you would rather be walked through the whole thing, our Hermes Agent course covers installation, messaging setup and writing skills.
Frequently asked questions
Is Hermes Agent free?
The software is free and published under the MIT license by Nous Research. Two costs remain: hosting, if you rent a server to keep it running, and the calls to whichever language model you plug in. Using a local model through Ollama on hardware you already own, the bill can drop to zero.
How is it different from ChatGPT or Claude?
ChatGPT and Claude are conversational assistants: you ask, they answer, and the next session starts with no memory of the last one. Hermes Agent is a self-hosted autonomous agent that takes a goal, chains actions on your machine and keeps its memory between sessions in a local database. It uses one of those models as its reasoning engine.
Does Hermes Agent run on Windows?
Not natively. You install WSL2, the Linux layer built into Windows, and run Hermes Agent from there. On Linux and macOS the install is direct. Android works through Termux, with limits on voice and media features.
Can you move from OpenClaw to Hermes Agent without reconfiguring everything?
Yes. The hermes claw migrate command, run during the initial setup, imports your settings, memories, skills and API keys from OpenClaw. Your model providers and messaging connections carry over.
Which language model should you run it on?
Any endpoint compatible with the OpenAI API works, including Nous Portal, OpenRouter, Anthropic or a local Ollama instance. The one hard requirement is a context window of at least 64,000 tokens, otherwise the agent's working memory does not fit on multi-step tasks.