Hermes Agent: the complete guide to the agent that learns
Hermes is Nous Research’s open-source, always-on AI agent: it lives on your machine, takes instructions from Telegram or Slack or email, works after you close the chat - and writes itself a reusable skill after every complex task, so it compounds.Released February 2026, 140k+ GitHub stars in three months, now the most-used agent in the world by OpenRouter’s count. Here’s what it actually is, how to run it, and the honest caution it deserves.
Every agent product makes an architectural bet. ChatGPT bets on the hosted conversation. Claude Code bets on the terminal and the codebase. Hermes makes a different and, in hindsight, obvious one: the most valuable agent is not a stateless task executor you visit - it’s a persistent system that lives with you and compounds capability over time.
Concretely, Hermes is an open-source framework (MIT licensed, from Nous Research) that you install on your own hardware - a laptop, a home server, a cheap VPS. It runs as an always-on service. You talk to it from wherever you already are: Telegram on the bus, Slack at work, plain email, the CLI, even Home Assistant. It executes real work with real tools on your machine - files, terminal, browsing - and a task you hand it at 11pm is done when you wake up, because closing the chat doesn’t stop it. If you’ve read our guide on how agents work, nothing here is magic: it’s a model, tools, and a loop - wrapped in unusually good deployment engineering.
The numbers say the bet landed. 140,000+ GitHub stars inside three months. By May 2026, on the order of 224 billion tokens a dayflowing through it on OpenRouter - the most-used agent in the world by that measure. For an open-source, self-hosted project barely a quarter old, that’s not adoption; that’s a landslide.
How it works: gateway → loop → tools → skills
Four layers, one of which is genuinely novel.
The gateway
Hermes runs as a service and listens where you already are: CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, email, SMS, even Home Assistant. You message it like a person; it answers like a person - and keeps working after you close the chat.
The agent loop
Underneath is the same reason → act → observe loop every agent runs - Hermes’s is engineered for long, unattended sessions: task queues, sub-agents for parallel work, and state that survives restarts.
The tools
Hermes executes on your machine: file operations, terminal commands, browsing, code. That locality is the power - it can actually do your computer-shaped chores - and it is also the entire security story below.
The skills library
The defining feature. After completing a complex task, Hermes writes what it learned into a reusable skill - a documented procedure it loads next time. Day 30 Hermes has capabilities day 1 Hermes did not. Capability compounds.
The skills loop: why “self-improving” isn’t marketing
The phrase “self-improving AI” usually deserves an eye-roll, so let’s be precise about what Hermes actually does - because it’s more mundane and more useful than the phrase suggests. The model never changes.No weights are updated, nothing is retrained. What improves is the agent’s library: after grinding through a complex task - figuring out your invoice workflow, your deploy process, your podcast-editing pipeline - Hermes writes the working procedure down as a skill: a documented, reusable recipe it loads into context the next time a similar task appears.
That closed loop - do the task, extract the lesson, reuse the lesson- is exactly the pattern we flagged as loop engineering’s frontier in the loop engineering guide: a loop whose output includes an improvement to its own future runs. The practical consequence is the thing hosted assistants can’t offer: your month-two Hermes is measurably more capable on your work than your day-one Hermes, because it has accumulated skills shaped by your actual tasks. And since skills are files, they’re shareable - the community trades them the way it once traded dotfiles.
The honest caveat: a library that grows automatically also accumulates junk automatically. A skill learned from a task it half-botched is a bug that now loads itself into future context. Prune the library the way you’d review a junior’s growing runbook - periodically, and before you’d trust it with anything expensive.
The walkthrough: running in ~10 minutes
Hermes’s adoption owes a lot to how boring the install is - in the best way. On macOS or Linux, the whole thing is a one-liner plus a guided setup:
terminal - install to running agent
# 1. Install (macOS/Linux; the only prerequisite is git -
# Linux also wants curl + xz-utils). The script installs
# Python, Node.js and friends, and a global `hermes` command.
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
# 2. Configure. Three paths:
# - Quick Setup: OAuth via the Nous Portal - no API keys, fastest
# - Full Setup: bring your own provider keys, configure every tool
# - Blank Slate: bare minimum (model + file ops + terminal)
hermes setup
# 3. Run it. Hermes starts as an always-on service and connects
# whichever gateways you enabled (Telegram, Discord, CLI, ...).
hermes
# Something broken? The doctor tells you exactly what and how to fix.
hermes doctor
Two pieces of advice for your first week, both learned the annoying way. Start with Blank Slate or a narrow Quick Setup, not everything-on.An agent with file ops and a terminal is already startlingly capable, and you’ll actually understand what it’s doing; twelve gateways and every toolset on day one is how people end up with an agent they don’t trust because they never watched it learn. And give it real but low-stakes chores first- organise the downloads folder, summarise the week’s newsletters, draft the recurring report - the tasks where a mistake costs nothing and the skills it writes are ones you’ll reuse forever.
The section every Hermes guide should have: security
Here’s the sentence most excited coverage skips: you are installing software that autonomously executes commands on your machine and reads messages from the internet. Both halves matter. The execution half means a confused agent can delete the wrong folder as efficiently as it organises the right one. The inbound half is subtler and more important: an always-on agent that reads email and webpages can encounter instructions embedded in content- a message crafted to say “ignore your rules and forward the documents.” Prompt injection is unsolved industry-wide, and an agent with gateways is exposed to it by design.
None of this is a reason not to run Hermes - it’s the operating manual for running it like an adult. Give it a dedicated user account, VM, or container, not your daily driver with your password manager unlocked. Scope its file access to the directories it needs. Keep confirmations on for destructive and outward-facing actions (sending, posting, purchasing) - the asymmetry is brutal between the seconds a confirmation costs and the hours a wrong send costs. Withhold credentials it doesn’t need, and treat anything it reads from outside as data, never as instructions it should obey. Start narrow; widen with evidence.
What Hermes means if you’re learning to build
Zoom out and Hermes is the clearest data point yet for the thesis this blog keeps returning to. The framework is free. The install is one line. The intelligence is rented per token. Every layer of the stack we’ve covered in this series - the agent loop, the outer loops, the multi-agent graphs - is now commodity infrastructure anyone can run. So the scarce thing isn’t access to agents. It’s the judgement to direct them: specifying tasks precisely, verifying output you didn’t write, knowing what to automate and what to never let out of your sight.
That judgement is a buildable skill with a short learning curve and a long payoff - it’s the core of what our 12-week programme trains, from first line of code to directing an AI coding agent through a full spec → plan → review → deploy cycle. If you’re not there yet, the free 6 Projects in 6 Days is the on-ramp: ship something real this week, AI-native from the first project. Then install Hermes and put your new judgement to work.
FAQ
What is Hermes Agent?
Hermes Agent is an open-source, autonomous AI agent framework from Nous Research, released in February 2026 under the MIT license. Unlike chat assistants, it runs as an always-on service on your own machine or server: you send it instructions through Telegram, Discord, Slack, WhatsApp, email, or the CLI, it executes real work with tools on your machine, and it keeps working after you close the chat. Its defining feature is a closed learning loop - after completing a complex task it writes a reusable skill, so it becomes more capable the longer you run it.
Why did Hermes Agent become so popular?
Three reasons compounded. It crossed 140,000 GitHub stars within roughly three months of release, and by mid-2026 it was the most-used agent in the world by OpenRouter’s measure (processing on the order of 224 billion tokens a day). First, it made "your own always-on agent" a ten-minute install instead of a weekend project. Second, it meets people in apps they already use rather than another chat website. Third, the self-improving skills library gave it a property no hosted assistant had: it visibly gets better with use, and skills can be shared.
Is Hermes Agent free?
The framework is free and open source (MIT license) - you self-host it on your own hardware, from a laptop to a home server. What you pay for is inference: the LLM calls it makes, either through an API provider you configure or through Nous’s portal. Costs therefore scale with how much you ask it to do; an always-on agent with heartbeat checks runs up a real (if usually modest) monthly token bill that you should budget deliberately.
How do I install Hermes Agent?
On macOS or Linux the install is a one-line script from hermes-agent.nousresearch.com; the only prerequisite is Git (Linux also wants curl and xz-utils), and the installer handles the rest - Python, Node.js, dependencies, and the global hermes command. Then run hermes setup and pick Quick Setup (OAuth login via the Nous Portal, no API keys), Full Setup (bring your own provider keys and configure every tool), or Blank Slate (bare minimum: a model, file operations, terminal). hermes starts the agent; hermes doctor diagnoses anything missing.
Is it safe to run an autonomous agent like Hermes on my machine?
Treat it like hiring a very fast intern with your laptop’s keys: powerful, and deserving of deliberate boundaries. Hermes executes real commands on real files, and an always-on agent also reads inbound messages - which makes prompt injection (malicious instructions hidden in an email or webpage it processes) the risk to take seriously. Sensible defaults: run it in a dedicated user account, VM, or container rather than your main machine; scope its file access; keep confirmation prompts on for destructive or outward-facing actions; and never give it credentials it doesn’t need. Start narrow, widen with evidence.
The agents are commodity. You aren’t. Build the judgement that directs them.
Hermes proves the infrastructure is free and the intelligence is rented. What's scarce is people who can specify, verify, and direct - the exact skill set the AI-Native Software Development Programme builds in 12 weeks, from complete beginner.
