Hermes Agent is an open-source AI agent from Nous Research that writes down what it learns and gets better at your work over time. It runs on your own laptop, a small server or a cloud sandbox, and you can talk to it from Telegram, Slack, WhatsApp and other chat apps. This guide explains how the Hermes Agent learning loop works, where it runs, which models it needs, how it differs from the Hermes models, and what the October 2026 funding news means for businesses.
The short version
What Hermes Agent is, and what it is not
Nous Research has made two things called Hermes, and they are easy to mix up. The Hermes models are language models, such as Hermes 3 and Hermes 4, that you can download and run. The Hermes Agent is a program that uses a language model to do work: run commands, read files, browse, send messages and remember. The agent does not need a Hermes model. It works with Claude, GPT, Gemini, Qwen, DeepSeek and many others.
| Hermes models | Hermes Agent | |
|---|---|---|
| What it is | Language models (for example Hermes 4, Hermes 4.3) | An agent program that uses a model to act |
| You get | Model weights | A CLI, a desktop app and a messaging gateway |
| Needs a Hermes model? | Not applicable | No; about 40 providers are supported |
| License | See each model card | MIT |
When Hermes Agent launched: sources disagree
There is no single launch date. The GitHub repository was created on 22 July 2025. The Nous Research releases page lists Hermes Agent on 25 February 2026, and most coverage, including The Next Web, says it was released in February. Some articles call 7 October 2026 the launch, but that date belongs to the funding round and the business product. The fairest reading: work started in public in mid-2025, and the official release was February 2026.
How the self-improving skill loop works
Most agents forget how they solved a problem once the chat ends. Hermes Agent keeps a written record. The Hermes Agent skills documentation describes the loop in plain terms.
Step 1: it saves what worked
After a task with several steps, the agent saves the procedure as a skill. It also saves fixes it found after hitting an error, and corrections you gave it. Small facts, like your time zone, go to memory instead.
Step 2: it reviews and tidies up
After a session, a review pass can suggest changes to skills. A curator pass merges overlapping skills. The agent prefers small edits over full rewrites, and skills should hold "lessons, not logs."
Step 3: it loads skills only when needed
Each skill is a SKILL.md file in ~/.hermes/skills/, compatible with the agentskills.io format. The agent first sees only short descriptions and opens the full file when a task needs it, which keeps the context small.
How you stay in control
By default, the agent writes skills without asking. You can change that. With approval turned on, every new or changed skill waits in a queue until you approve or reject it, and you can view the diff first. You can lock a skill so the agent cannot change it, or simply edit or delete the file. For the manual way of creating skills from a folder or URL, see our guide to the /learn command.
Where Hermes Agent runs
The README lists six built-in places where the agent can run its commands, plus plugins for more:
- Local: your own machine on Linux, macOS or WSL2, with a desktop installer for macOS and Windows.
- Docker or SSH: a sandbox on your machine or a remote server.
- Modal, Daytona or Vercel Sandbox: cloud sandboxes. Modal and Daytona can sleep when idle and wake up with their state intact.
The README says it can run on a $5 VPS or a GPU cluster. Installing takes one command, and the quickstart recommends getting a plain chat working before adding anything else.
A sensible first setup
- 1Install and pick a provider. Run the installer, then choose a model with the setup wizard.
- 2Test a plain chat. Confirm that a session resumes after you close it.
- 3Move commands into a sandbox. Switch the terminal backend to Docker or SSH before you give the agent real work.
- 4Turn on skill approval. For the first weeks, review every skill it wants to write.
Messaging channels
One gateway process connects the agent to your chat apps. The Hermes Agent quickstart lists Telegram, Discord, Slack, WhatsApp, Signal, email, Microsoft Teams and Home Assistant (as a plugin). That means you can ask your agent for a report from the same app your team already uses. For a phone-first way to manage it, read our post on Hermes Workspace Mobile.
Which models to use with Hermes Agent
The one hard rule: the model needs at least 64,000 tokens of context. Smaller windows are rejected at startup. The quickstart notes that hosted models from Anthropic, OpenAI, Google, Qwen and DeepSeek meet this easily. It suggests these paths:
- 1Nous Portal: the easiest start. One subscription covers more than 300 models.
- 2Your existing provider: OpenAI, Anthropic, OpenRouter, Google AI Studio, xAI and about 40 others. You can switch at any time.
- 3Local models: through Ollama or llama.cpp, as long as you set the context to at least 65,536 tokens.
Nous Research does not name one best model. Our advice: start with a strong hosted model so the skills it writes are good, then test cheaper or local models on tasks it already has skills for.
License, funding and the business version
Hermes Agent is released under the MIT license, so you can use, change and ship it, including in commercial products. On 7 October 2026, Nous Research announced a $90 million round. The Wall Street Journal reported a $1.5 billion valuation; according to The Next Web, Nous did not confirm that figure. Named investors include Nvidia, Microsoft's M12, Samsung, Robot Ventures, Union Square Ventures, Y Combinator and Menlo Ventures.
The money goes into Hermes for Businesses, which Nous says will let companies pick models by price, keep their AI use private and own their company knowledge. Nous also claims more than 24 million clones since February and about 2.5% of global token usage. These are company figures, not independent ones. Prices for the business tiers were reported differently by different outlets, so check with Nous directly.
Is Hermes Agent right for your team?
- Good fit: you want to keep data on your own servers, choose your own model, and build up a library of team skills over time.
- Good fit: your team already lives in Slack, Telegram or Teams and wants the agent there.
- Think twice: nobody on the team can look after a server, update the agent and review what it writes.
- Think twice: you need a fully managed product with a support contract today.
If you prefer a managed agent, compare it with OpenAI Dots.
How Halmob helps
A self-improving agent is only useful if someone checks what it learns. Our agent platform consultancy helps businesses choose, set up and roll out Hermes Agent, OpenClaw, Pi Durable, Meta Muse, OpenAI Dots and Claude Duet, including skill approval and model choice. We also build a software factory: an agent pipeline that turns a spec into reviewed, tested pull requests in your repo. Get in touch to plan your setup.