Hermes Agent builds its own skills the more you use it

The third announcement, this time about the learning loop

On August 17, Nous Research introduced bot mode for Hermes Desktop, which splits agent profiles by bot, and followed up on August 20 with a dedicated remote computer and the freedom to fork and self-host. This latest repository description is the next installment in that series, laying out the learning loop that serves as the agent's core engine along with how to install and integrate it.

Runs anywhere

According to the company, Hermes Agent can run on a $5 VPS, a GPU cluster, or serverless infrastructure that costs almost nothing when idle. It isn't tied to your laptop — you can message it on Telegram while it handles a task on a cloud virtual machine. There are two ways to run it: launch the TUI in a terminal with the hermes command, or turn on the gateway and talk to it through Telegram, Discord, Slack, WhatsApp, Signal, or email. The two interfaces share most of the same slash commands.

Model choice is flexible too. You can switch between multiple providers — including Nous Portal, OpenRouter, OpenAI, and your own endpoints — with a single hermes model command, and the company says this requires no code changes and creates no lock-in.

Getting it running

Installation paths differ depending on your operating system.

Environment	Installation method	Install path
macOS/Linux	One-line script	~/.hermes
Windows (native)	PowerShell one-liner, no WSL needed	%LOCALAPPDATA%\hermes
WSL2	Same commands as Linux	~/.hermes
Android (Termux)	Manual install following the Termux guide	Separate extended build

On Windows, the CLI, gateway, TUI, and tools all run natively without WSL. The installer automatically sets up uv, Python 3.11, Node.js, ripgrep, ffmpeg, and even a roughly 45MB portable Git Bash (MinGit). It's installed in an isolated location at %LOCALAPPDATA%\hermes\git without needing admin rights, so it doesn't conflict with any Git already on the system, according to the company. Antivirus software sometimes flags and quarantines uv.exe — the company explains this is a false positive caused by uv, the Python package manager, being distributed unsigned. On Android, the full extended build pulls in voice-related dependencies that don't work on Android, so a separate build trimmed down for Termux is installed instead.

For existing OpenClaw users, the setup wizard (hermes setup) automatically detects the ~/.openclaw folder and offers to migrate it. It can carry over your OpenClaw settings, memory, skills, and API keys as-is. You can check the options with hermes claw migrate --help, or go through it interactively via the openclaw-migration skill.

Five keys, one subscription

Normally, connecting multiple tools means issuing separate API keys for the base model, web search, image generation, TTS, and cloud browser access. Nous Portal bundles all five into a single subscription.

Feature	Standalone use	Nous Portal
Model	Separate API key needed	Bundled into one subscription
Web search	Separate API key needed	Bundled into one subscription
Image generation	Separate API key needed	Bundled into one subscription
TTS	Separate API key needed	Bundled into one subscription
Cloud browser	Separate API key needed	Bundled into one subscription

Signing in with OAuth sets Nous as the provider and turns on the Tool Gateway. You can check connection status anytime with hermes portal info, and you're still free to use your own API key for individual tools. Since the gateway is set up separately for each backend, it's not all-or-nothing.

Editor's view

Counting up Nous Research's announcements around Hermes Agent over the past month gets you four: bot mode, the remote computer, freedom to fork, and now this learning loop. Taken separately, they read like a feature list. Put them together, and a clear direction emerges. Rather than pushing users toward one model and one app the way OpenAI or Anthropic does, Nous seems to be building a modular agent that works with any model and runs on any infrastructure. Both the design that lets you switch providers with a single hermes model command and the gateway's per-backend setup point in the same direction.

Anyone who's actually put an agent framework this size into real use runs into the same wall every time. It works fine for the first few days, but once conversations pile up, you inevitably hit a point where you have to re-explain earlier context. The learning loop Hermes is touting — building its own skills and pulling up past conversations on demand — is aimed squarely at that problem. That said, long-term-memory designs like this only prove themselves after weeks of actual use, so it's too early to judge their effectiveness from the launch description alone.

For teams here, the migration path from OpenClaw is worth watching closely. Being able to bring over settings, memory, skills, and API keys with a single command means that teams who've already built workflows around OpenClaw may find the switching cost lower than expected. On the other hand, teams adopting an agent for the first time will likely find that consolidating five separate APIs under one Nous Portal subscription noticeably cuts down setup time.

Over the coming weeks, we should start seeing real cases showing how many skills this learning loop actually generates and how users go about verifying their quality. For now, all we have is the company's own description — the next chapter belongs to the people who've actually tried it.