Hermes Agent
Hermes Agent is a free, open-source AI assistant that lives on your own computer or server, remembers what it learns, and gets better at your tasks the longer you use it. Use it for always-on personal automation you can reach from Telegram or WhatsApp. For sit-down coding in a codebase, Claude Code is still the stronger tool.
What it is#
Who makes it: Nous Research, an open-model AI lab known for its own "Hermes" family of language models. The agent shares the name but is a separate product.
What it is: an agent "harness", meaning the program that wraps around an AI model and gives it hands (tools), a memory and a way to reach you. You bring the brain: Claude, GPT, Gemini, DeepSeek, Kimi or a model running on your own machine.
Released: February 2026, under the MIT license, so it is free to use, copy and change, even commercially.
Tagline: "The agent that grows with you." Its big idea is a built-in learning loop: it writes down reusable how-tos (skills) from work it has done and refines them each time.
Where it lives: wherever you put it. A $5-a-month cloud server, your laptop, a GPU machine, or pay-when-used cloud sandboxes that sleep when idle. You talk to it from the terminal, a desktop app, or a chat app on your phone.
How big it is#
| Signal | Figure |
|---|---|
| GitHub stars | About 252,000 (26 Oct 2026) |
| Downloads | Cloned more than 24 million times since February |
| Funding | $90M Series B at a $1.5B valuation, led by Robot Ventures with NVIDIA, Microsoft M12, Samsung and USV |
| Latest major release | v0.20.0: real-time voice, wake words, agent-to-agent links, signed webhooks, longer autonomous runs |
| Daily usage | Passed OpenClaw in daily token volume in mid-2026 |
It grew fast partly by timing: in April 2026, when Anthropic stopped flat-rate Claude Pro and Max plans being used inside the rival agent OpenClaw, Nous pitched Hermes to those users, with a one-command import of their OpenClaw setup.
How it works#
Think of Hermes as one assistant sitting in the middle. You reach it from any screen, it thinks with whichever AI model you plug in, and it keeps its own notebook.
The agent loop is the core: it reads your request, decides which tool to use, runs it, checks the result and repeats until the job is done. Everything else serves that loop.
| Part | What it does in plain words |
|---|---|
| Gateway | One background program that connects Hermes to 20+ chat apps, so a Telegram message and a terminal command reach the same agent |
| Memory | Plain text files it keeps about facts (MEMORY.md) and about you (USER.md), plus a search over past conversations that it summarises when needed |
| Skills | Step-by-step how-tos saved as files. It writes new ones from experience and rewrites them when it finds a better way. They follow the open agentskills.io format, so they also work in other agents |
| SOUL.md | A file that sets its default personality and tone |
| Context files | Project notes it reads at the start of every conversation, like Claude Code's CLAUDE.md |
| Subagents | Copies of itself it spins up to work on several things in parallel |
| Profiles | Separate Hermes instances, each with its own memory, skills and settings, e.g. one for work and one for personal |
| Terminal backends | Where its commands actually run: your machine, a Docker container, a remote server over SSH, or cloud sandboxes (Modal, Daytona) that sleep when idle |
The learning loop#
This is what makes Hermes different from most agents, which forget everything when the chat ends.
Say you ask it to compile a weekly report from three websites. The first time, it works out the steps by trial and error. Afterwards, a built-in nudge asks it what was worth keeping, so it saves a "weekly report" skill. Next week it loads that skill, skips the trial and error, and updates the skill if a site has changed. Over months, your Hermes collects dozens of these, tuned to how you work.
Install and setup#
Easiest: download the Hermes Desktop installer for Windows or Mac from hermes-agent.nousresearch.com. It installs both the app and the command-line tool.
Command line only:
| System | Command |
|---|---|
| Linux, macOS, WSL2, Android (Termux) | curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash |
| Windows (native, early beta) | iex (irm https://hermes-agent.nousresearch.com/install.ps1) |
On Windows, WSL2 is the most tested route; native Windows is still early beta.
First run, step by step:
hermes setup --portalsigns you in to Nous Portal and gives you a model plus web search, image generation, voice and a browser in one go. Or run plainhermes setupto plug in your own API keys instead.hermesstarts chatting in the terminal.hermes gateway setupthenhermes gateway startconnects it to Telegram, WhatsApp or another chat app, so you can message it from your phone.Coming from OpenClaw?
hermes claw migrateimports your memories, skills, API keys and chat settings. Setup also spots an existing OpenClaw folder and offers this automatically.
Everything lives in a ~/.hermes folder on your machine: config, memory, skills and logs, all readable plain files.
Which model to run it on#
Hermes works with 200+ models through providers including Nous Portal, OpenRouter, Anthropic, OpenAI, Google, DeepSeek, Kimi, MiniMax, local Ollama, or any OpenAI-compatible server.
| Choice | Good for | Note |
|---|---|---|
| Claude Opus 5.5 or Sonnet 5.5 (API key) | Best reasoning on long, messy tasks | Use a pay-as-you-go API key; flat-rate Claude plans have been restricted for third-party agents |
| DeepSeek, Kimi or other budget models | Routine skills and scheduled jobs | Far cheaper; caching makes repeated runs very cheap |
| Nous Portal | One login for 300+ models plus four tools | Free tier, or Plus at about $20/month with tools bundled |
| Local model (Ollama) | Privacy, zero API bill | Needs a strong machine; weaker on complex multi-step work |
A common pattern: a strong model for planning, a cheap one for routine jobs.
Key features, with examples#
| Feature | What it means | Example |
|---|---|---|
| Chat from anywhere | Same agent across 20+ apps, including Telegram, WhatsApp, Slack, Discord, Signal, Teams, email and SMS | Message it from your phone on the train: "check my server logs and tell me if anything broke overnight" |
| Scheduled jobs (cron) | Runs tasks on a timer and sends results to any chat app | Every morning at 8: summarise AI news and post it to your Telegram |
| Skills | Saved how-tos it writes and improves; browse with /skills, install community ones with hermes skills install | After you show it once how to post your meme page's weekly roundup, it saves the steps and repeats them |
| Persistent memory | Remembers projects, preferences and people across sessions and restarts | It already knows your brand colours when you ask for a social graphic next month |
| Subagents | Spins up parallel copies for separate workstreams | Research five competitors at once, one subagent each |
| execute_code | Writes a small script to do many tool calls in one go, saving time and tokens | Pull 200 rows from an API, clean them and save a CSV in one step |
| Browser and computer use | Searches, reads pages, clicks through sites | Fill a supplier's web form with data from a spreadsheet |
| MCP | Connects any MCP server (Notion, GitHub, databases) | Read your Notion projects page and update statuses |
| Voice mode | Real-time voice in terminal, Telegram and Discord, with wake words since v0.20.0 | Talk to it hands-free while you cook |
| Webhooks | Other apps can trigger it, and it can send signed alerts out | A new form submission triggers Hermes to draft a reply |
| Hermes Workspace | Community web dashboard for chat, memory, skills, jobs and terminal in one browser tab | Browse and edit what it remembers without opening files |
Useful commands: /new starts a fresh chat, /compress shrinks a long conversation, /usage shows token spend, /stop or Ctrl+C interrupts work, hermes sessions export saves history to a file.
Hermes vs Claude Code vs Codex vs OpenClaw#
These overlap but are built for different jobs. Hermes is agent-first, OpenClaw is gateway-first, and Claude Code and Codex are coding-first.
| Hermes Agent | Claude Code | Codex | OpenClaw | |
|---|---|---|---|---|
| Made by | Nous Research | Anthropic | OpenAI | Open source community |
| Main job | Persistent all-round assistant | Coding in a repository | Coding in a repository | Always-on assistant hub |
| Models | Almost any, including local | Claude only | OpenAI models | Almost any, including local |
| Memory | Built-in learning loop, skills it writes itself | CLAUDE.md, AGENTS.md, auto memory | AGENTS.md, project context | Workspace files plus search |
| Reach it from | Terminal, desktop app, 20+ chat apps | Terminal, IDEs, desktop, web, Slack | Terminal, IDE, web | Many chat apps via plugins |
| Scheduling | Built-in cron | Routines and scheduled tasks | Cloud tasks | Gateway automation |
| Cost | Free software; pay for server and model | Claude plan or API | ChatGPT plan or API | Free software; pay for server and model |
| License | MIT | Commercial | Commercial app, open CLI | MIT |
How to choose:
Writing, debugging and refactoring code in a project: Claude Code (or Codex if you are on OpenAI). They have the best code awareness, tests, Git and IDE support.
An assistant that runs 24/7, remembers you, works from your phone and handles mixed tasks (research, reports, monitoring, messages): Hermes.
You already run OpenClaw with lots of its 5,000+ community skills: stay, or migrate with one command if you want Hermes's learning loop.
Many people run both: Claude Code for sit-down coding sessions, Hermes for one-off code jobs and everything else on the move.
Costs#
The software is free. You pay for three things:
| Cost | Typical range | Note |
|---|---|---|
| Server | $4 to $25/month | 1 GB RAM and 1 CPU is enough when the model runs in the cloud; skip it entirely by running on your own laptop |
| AI model usage | $2 to $60/month | Biggest variable. Premium models can cost about 30x more than budget ones for the same work |
| Optional tools | A few dollars, or Nous Portal Plus at about $20/month | Web search, browser, image generation, voice |
A light personal setup runs $5 to $8 a month; heavy use on premium models reaches $80+. Managed hosting (for example Hostinger) starts around $6/month if you do not want to run a server.
Two cost traps: chatting through Telegram or other apps uses roughly 15,000 to 20,000 tokens per turn, versus 6,000 to 8,000 in the terminal, because of the extra context; and hourly-billed cloud servers left on 24/7 add up fast.
Limitations#
Not the best coder. For deep work inside a codebase, Claude Code and Codex are stronger.
Setup and upkeep are on you. Models, keys, server, updates and security settings are your job unless you pay for managed hosting.
Young project. Released February 2026 with frequent releases; features and commands change quickly.
Native Windows is early beta. Use WSL2 or the desktop app.
Quality depends on the model. Cheap or local models stumble on long multi-step tasks, and a mistake early on can derail the whole run.
Memory can go wrong. It decides itself what to save, so it can keep something wrong or private and reuse it later.
Costs can creep. Long autonomous runs (now up to 500 tool steps by default) and chat-app usage burn tokens.
Security#
Hermes can run commands, read files, browse, and message people with whatever access you give it, so treat it like a server, not a chat app.
The only real wall is your operating system. Nous's own security policy calls Hermes a single-person agent whose true safety boundary is OS-level isolation. By default it runs commands on your machine with your permissions.
Run it in a box. Use the Docker or a remote backend, not your main laptop's local shell, for anything serious.
Review skills before installing. Skills and plugins run with full agent access. The built-in Skills Guard scanner helps but is not a guarantee.
Lock down who can talk to it. Use allowlists and pairing so only you can message your agent.
Keep approvals on for risky terminal commands.
Audit its memory now and then in
~/.hermes, and delete what it should not keep.Known bugs happen. Public vulnerabilities have included a path-traversal flaw (CVE-2026-7396) that let it read files outside its folder; keep it updated.