Hermes Agent is an open source, self hosted AI agent from Nous Research that runs persistently on your own machine, remembers what it learns across sessions, and writes reusable skills so it gets more capable the longer it runs. Released in February 2026 under an MIT license, it shipped an official desktop app on June 2 2026 and has become one of the fastest growing agent projects of the year.

What is Hermes Agent?

Hermes is not a coding copilot tied to an editor or a chatbot wrapped around a single API. It is a persistent autonomous agent that lives on your server or laptop and connects to whatever large language model you choose, whether that is Claude, GPT, Gemini, or a local model run through Ollama. It ships with more than 40 built in tools, including web search, browser automation, file management, and code execution.

Reach is a big part of the pitch. Hermes connects to messaging platforms like Telegram, Discord, Slack, WhatsApp, and Signal, plus a command line gateway, so you can talk to it wherever you already are. Installation is a single curl command on Linux, macOS, or WSL2, with no manual prerequisites.

What makes it different from other agents?

The defining feature is that Hermes attacks what its creators call agent amnesia. Most agents forget everything when a session ends. After Hermes completes a hard task, it distills a reusable skill document and stores it in persistent memory, and a three layer memory system holds your preferences, projects, and context across restarts. The skills it writes are searchable, shareable, and compatible with the agentskills.io open standard, so the agent compounds its own knowledge over time.

In practice that means the second time you hand it a similar problem, it recalls the relevant skill and solves it faster. The longer it runs, the better it knows your environment, which is the opposite of starting from a blank slate every conversation.

Who is it for and what are the catches?

Hermes targets developers and tinkerers who want a self hosted agent with no telemetry, no tracking, and no cloud lock in, where every line of code is auditable and all data stays local. You still need to pay for an LLM provider's API or run a model on your own hardware, so it is free software with a compute bill attached.

There are rough edges. The Windows installer is not code signed, so Windows SmartScreen will flag it on first launch until you click through. Even so, the project crossed well over 90,000 GitHub stars within months of release, a clear signal of momentum in a crowded open agent field.

Check out what else is trending at GitHub Trending