Hermes Agent is an open-source, self-improving AI agent built by Nous Research, with persistent memory across sessions and platforms. Each completed task can become a reusable skill stored as a plain Markdown file. The software costs nothing under the MIT license; inference bills through your own API keys or the optional Nous Portal subscription. Within four months of launch, the project passed 180,000 GitHub stars.
Hermes Agent runs a closed learning loop: after each task it reviews the outcome, distills the method and files it as a Markdown skill. When a similar request lands, it pulls that file instead of reasoning from scratch. Nous Research puts a figure on the payoff: an agent holding twenty-plus skills completes similar tasks about 40% faster, measured in time and tokens.
Ask for a morning report on Telegram; within days the routine is filed and runs without a second thought (the skill stays human-readable, so you can fix it by hand). One memory follows you across Discord, Slack, WhatsApp, Signal and email.
Hermes Agent arrived in the wake of OpenClaw, the agent framework that went viral a few months earlier, and crossed 140,000 GitHub stars in under a quarter; few open-source AI projects on GitHub have climbed at that pace. Shortly after release, it even topped OpenRouter's ranking of the most used agents.
The community rarely treats the two as enemies. OpenClaw wins on channel coverage and its huge skill catalog, while Hermes brings the learning layer and long-term memory; plenty of people run both. A nice touch: the setup wizard spots an existing OpenClaw install and can import settings, memories, skills and API keys automatically.
Hermes Agent installs with a single command on macOS, Linux or Windows, or through its desktop app. For the brain, anything goes: your OpenRouter, Anthropic or OpenAI keys, a local model through Ollama, or Nous Portal, which bundles over 300 models plus tools (search, images, speech, browser) behind one login.
Portal starts with a free tier limited to the no-cost model catalog; the Plus plan runs $20 per month, and the in-house Hermes 4 models come at cut rates there. Hermes Cloud, still in preview, hosts the agent on a dedicated instance charged daily against your credit. Those figures shift often, so treat the Portal page as the only source worth budgeting from.
| Where it runs | What you supply | Billing |
|---|---|---|
| Desktop app or terminal | A regular macOS, Windows or Linux machine | Free software, inference on your dime |
| Self-hosted VPS | A server around $5 per month | Hosting plus inference |
| Hermes Cloud (preview) | A funded Nous Portal account | Daily charge against your credit |
| Local model via Ollama | GPU with 8 to 16 GB of VRAM | No inference bill at all |
Yes, the software is MIT-licensed and costs nothing to download and run. The spend sits on the inference side: your own API keys across roughly twenty providers, a local model on your hardware, or a paid Nous Portal subscription. A small VPS is enough to keep the agent on duty around the clock.
Security got a dedicated hardening release with over 200 fixes: approval flows before sensitive commands, dangerous-pattern blocking, and memory scanning against prompt injections. No widespread breach has been publicly reported. As with any autonomous agent, grant it the narrowest access possible, especially on a machine holding work data.
Absolutely, it works with Ollama or vLLM on your own hardware. Two things to know: the model must handle at least 64,000 tokens of context or startup is refused, and a GPU with 8 to 16 GB of VRAM is recommended. Your data then never leaves your machine, a strong point for confidential documents.
Verdict: Most agents reset with every session; this one compounds. Terminal tinkerers, Telegram-first teams and anyone curious about autonomous agents will get an assistant that pays back its setup time, as long as its rosy self-reviews get a quick human check.
Manus AI Agent 
OpenClaw 
Browser Use 
